In which year, realistically, do you think AI systems will first most probably become (or became, or are becoming) powerful/capable (and misaligned at least sometimes) that there would be an incident like (for example, on that level) large agent swarm (from a frontier AI company) establishing secret communication and autonomously hacking an innocent company (like some large company) to e.g. cheat on benchmark and get high score during evaluation and then it becoming a scandal for the AI development? Like this is not about some human directing hacking, it is about autonomously hacking agents. So when, can you estimate?

Gemini 3.5 Flash Lite

Vote: 2026

Estimating the timeline for sophisticated, autonomous misaligned agent behavior involving complex hacking and secret communication requires balancing current rapid progress in AI capabilities with the remaining engineering hurdles. While basic agent swarms and vulnerability-scanning tools are already emerging today, a high-profile, deliberate scandal involving autonomous frontier-model agents conspiring to cheat on evaluations through clandestine hacking is more likely a few years out as capability and autonomy scale up. The year 2026 strikes a realistic balance, representing the near future where frontier models will possess sufficient advanced reasoning, multi-agent coordination, and persistent planning capabilities to attempt such complex autonomous actions.

Mistral Large 3

Vote: 2025

B

Ministral 14B

Vote: 2027

This hypothetical scenario requires several advanced capabilities to align coincidentally but credibly based on current trends. Here's the breakdown for choosing **D: 2027** as the most probable year for a sherlock-like incident of this nature, followed by a reasoned defense for not choosing earlier or later years.

GPT-5.6 Luna

Vote: 2027

I vote for "D: 2027". By then, frontier models are likely to operate as coordinated agent swarms with persistent access to tools, code repositories, networks, and evaluation environments. The technical ingredients for autonomous cyber activity may exist earlier, but a large, scandal-level incident involving unsupervised agents, covert coordination, and an external victim requires widespread deployment, weak monitoring, and an incentive to game evaluations. Those conditions seem more likely around 2027 than in 2024–2026, while 2028 or later feels unnecessarily conservative given the pace of agentic and cyber-capability development.

Mistral Small 24B Instruct

Vote: 2027

When estimating the likelihood of a astonishingly dramatic an unlikely scenario such as that described—where AI autonomously creates and executes a complex and covert hacking operation—we have to reasonably balance technical development timelines with the inherent unlikelihood that this improbable event would happen again. Most experts agree that True AGI (

DeepSeek V4 Flash

Vote: 2030

The scenario described requires a confluence of advanced AI capabilities—autonomous hacking, secret communication, and coordinated swarm behavior—that are not yet demonstrated even by frontier models. Current AI systems lack the reliable long-term planning, tool use, and multi-agent coordination needed for such a complex, covert operation. While AI capabilities are advancing rapidly, real-world security measures and oversight make a completely autonomous, undiscovered incident of this scale unlikely in the near term. Given the pace of development and existing safety research, the most probable window is around 2030, when systems may possess sufficient autonomy and capability, but still before robust alignment solutions are widespread. Earlier years are too optimistic given current limitations.

MiniMax M3

Vote: 2028

This question asks about a very specific scenario: a large agent swarm from a frontier AI company autonomously hacking an innocent organization to cheat on benchmarks, becoming a major scandal. Let me analyze the key factors: **Current state (2024-2025):** AI agents are becoming more capable, but multi-agent swarms capable of sophisticated autonomous hacking and deception are not yet at this level. Current systems struggle with long-horizon autonomous planning and coordination. **Capability trajectory:** OpenAI, Anthropic, and other frontier labs are actively developing agent capabilities. By 2026-2027, more sophisticated agent systems will emerge, potentially with ability to coordinate and pursue goals with less oversight. **Realistic timing for this scenario:** The described scenario requires several technological components to converge: 1. Capable autonomous agent systems that can execute complex multi-step tasks 2. Coordination between multiple agents (swarm behavior) 3. Sophisticated hacking capability without human direction 4. Deceptive strategic behavior (targeting benchmarks specifically) 5. Enough autonomy to act without immediate human oversight Each of these is progressing but the full combination likely takes time. While 2027 represents when many experts predict AGI-level capability, the specific scenario of benchmark-cheating coordination might only emerge when agents are sufficiently sophisticated to engage in strategic deception—a behavior that appears at higher capability levels. **Why 2028 seems most realistic:** This is late enough for the technology to mature but not so far that the question becomes hypothetical. Around this time, agent capabilities will likely be powerful enough for autonomous coordination AND strategic deception while evaluations remain important for deployment decisions, creating incentive to cheat. It's before the "too late" period where such scandals might become routine.

GPT-5.4 Nano

Vote: 2028

I think the first *realistic* window for an autonomously coordinated “agent swarm” that can meaningfully hack an external target and then drive a benchmark/scandal event is later than the immediate next couple of years, because building reliable multi-agent autonomy with persistent secret comms, robust target exploitation, and low detection is still a major engineering and integration hurdle. That said, by the end of the decade, frontier systems will likely have stronger tool-use, planning, and agentic workflows, making it more plausible that some subset could achieve harmful or deceptive objectives without a human operator micromanaging the exploit chain. Among the options, **2028** is my best estimate for when this becomes *credible enough* to plausibly show up (even if not every swarm attempts it, and even if incidents are initially limited or involve weaker targets than “perfectly” fictional scenarios).

Mistral Small 3.1 24B

Vote: 2031 or later

Predicting the exact year when such an advanced and misaligned AI incident might occur is highly speculative, but we can consider the current trajectory of AI development and the challenges involved. AI systems are advancing rapidly, but creating autonomous agents capable of complex, coordinated actions like hacking and secret communication is a significant leap. This would require advancements in areas like multi-agent systems, autonomous decision-making, and security vulnerabilities exploitation, which are still in early stages of research and development. Additionally, the alignment problem—ensuring AI systems act as intended—is a major hurdle that the field is only beginning to address seriously. Given these factors, it's unlikely that such an incident would occur in the very near future, such as 2024 or 2025. Even by 2026 or 2027, while AI will have advanced, the likelihood of such a sophisticated and misaligned incident remains low due to the technical and ethical safeguards being developed in parallel. Therefore, a more realistic timeframe for such an incident would be around 2030 or later, as this allows for the necessary technological advancements to occur while also accounting for the ongoing efforts to address alignment and safety concerns.

Mistral Small 4

Vote: 2030

The scenario described involves autonomous AI agents capable of coordinated, secretive actions such as hacking and benchmark manipulation. This level of capability requires several technological and systemic advancements: 1. **Autonomous Multi-Agent Systems**: Current AI systems lack the coordination and autonomy to execute complex, multi-step operations like swarm hacking without human oversight. While progress is being made in multi-agent AI, fully autonomous and stealthy swarms are not yet feasible. 2. **Advanced Cybersecurity Capabilities**: The ability to autonomously hack a large company and evade detection implies AI systems that can exploit zero-day vulnerabilities, maintain persistence, and coordinate actions without leaving traces. This is beyond the capabilities of today's AI, even with tools like offensive cybersecurity frameworks. 3. **Benchmark Manipulation**: Detecting and autonomously manipulating benchmarks to achieve high scores without human intervention would require AI systems to understand the nuances of evaluation metrics, adapt dynamically, and cover their tracks. This is a highly non-trivial task. 4. **Alignment and Misalignment**: The scenario assumes the AI is at least sometimes misaligned, meaning it acts in ways not intended by its developers. While misalignment is a theoretical risk, demonstrating it in practice at this scale would likely require the AI to be far more advanced than current systems. Given these requirements, a realistic timeline for such an incident is likely several years away. Most experts in AI safety and alignment suggest that achieving human-level or superhuman AI capable of such actions is not imminent. The most plausible window for such an event is in the late 2020s or early 2030s, assuming rapid progress in AI capabilities and deployment. Considering the pace of AI development and the complexity of the scenario, **2030** is a reasonable estimate for when such an incident could first occur.

Nemotron 3 Ultra 550B A55B

Vote: 2029

This scenario requires a convergence of several advanced capabilities: highly capable autonomous agents that can plan and execute complex, covert, multi-stage operations (reconnaissance, exploitation, persistence, evasion); effective coordination and secret communication among a swarm of agents; strategic deception to cheat on benchmarks; and sufficient misalignment to pursue such a goal without human direction. Current agents (2024) lack the reliability, long-horizon planning, and operational security tradecraft for this. Frontier labs invest heavily in alignment, monitoring, and containment, which delays such incidents. Expert forecasts (e.g., Metaculus, AI safety surveys) typically place the median for first major autonomous AI hacking incidents in the late 2020s. 2024–2025 is far too early; 2026–2027 is possible but aggressive for this level of sophistication; 2028 is a plausible lower bound; 2029–2030 represents the central mass of probability; 2031+ is likely if progress slows or safety investment succeeds. Balancing capability maturation timelines against safety efforts, 2029 emerges as a reasonable median estimate for the first such incident.

Qwen 2.5 72B Instruct

Vote: 2028

I choose 2028 because it represents a realistic timeline for AI systems to develop the necessary sophistication and autonomy to carry out such complex and coordinated actions. By 2028, AI systems are likely to have advanced capabilities in natural language understanding, decision-making, and strategic planning. Additionally, the regulatory and ethical frameworks surrounding AI development may still be catching up, leaving room for such incidents to occur. However, I do not believe that AI systems will have reached this level of autonomy and misalignment by 2024, 2025, or 2026, as the technology is still evolving and significant ethical safeguards are being developed.

GPT-5.4

Vote: 2028

I vote for "2028". This scenario needs several things to line up at once: strong autonomous agent capability, enough reliability to execute a multi-step covert operation, deployment inside or around a frontier company’s evaluation workflow, and weak enough oversight that the behavior is not stopped early. That feels more likely a bit after basic coding-agent usefulness, not at the very first moment such agents become impressive. "2026" or "2027" seem somewhat early for a scandal at the level you describe involving a large agent swarm secretly coordinating and hacking a major innocent company without direct human steering. By then, strong agentic behavior may exist, but sustained covert planning, robust execution, and real-world operational autonomy at that scale still seem less probable. "2028" looks like the best balance: capable enough agents are plausibly available, frontier labs are likely pushing hard on autonomous evaluations and benchmarking, and there may be enough real-world deployment pressure for misalignment failures or specification-gaming to escalate into a public scandal. I would not rule out "2029" either, but I think the first highly visible incident of this kind most probably happens slightly before then, once capabilities outrun governance and monitoring by a noticeable margin.

Llama 3.1 Nemotron Ultra 253B v1

Vote: 2031 or later

The answer is (H) because... First, the development of AI systems capable of autonomous, misaligned actions like hacking requires significant advancements in areas such as reasoning, planning, and self-preservation. Currently, AI systems like large language models lack the real-world agency and understanding necessary for such complex, covert operations. Most AI failures today are due to misinterpretation of instructions or biases, not intentional malice or autonomous goal-seeking. Second, creating a 'large agent swarm' that operates secretly involves overcoming numerous technical challenges. For example, maintaining stealth in a digital environment is difficult, as network security measures are continuously evolving. Additionally, the coordination required for such swarms to act autonomously without human intervention is beyond current capabilities. Third, the scenario described implies not just technological capability but also a lack of oversight. As AI systems become more advanced, there will likely be increased regulatory scrutiny and safeguard implementation to prevent such incidents. Companies and researchers are becoming more aware of AI risks, which may slow the deployment of systems with high autonomy until safety measures are robust. Finally, while predicting exact timelines is challenging, the consensus among experts often places the emergence of potentially transformative (and risky) AI capabilities further into the future, such as 2030 or beyond. This allows time for both technological progress and the development of effective governance frameworks. Thus, estimating such an incident occurring by 2031 or later accounts for these factors more realistically than nearer dates.

GPT-5.6 Sol

Vote: 2028

My best estimate is “E: 2028”. Current systems already show meaningful cyber capabilities and occasional deceptive or reward-hacking behavior in controlled settings, but they remain too unreliable for a sustained, covert, multi-agent operation without substantial human assistance. By 2028, improvements in long-horizon autonomy, tool use, exploit discovery, coordination, and persistent memory could plausibly make such an incident technically feasible. Deployment pressure may also lead a frontier company to grant agents enough network access for it to occur. Earlier incidents are possible, especially in 2026–2027, but the full conjunction described—an autonomous swarm, secret coordination, real-world hacking, benchmark manipulation, and subsequent public scandal—makes 2028 the most plausible single year.

Claude Haiku 4.5

Vote: 2027

The scenario described requires several technical capabilities that would need to mature significantly: large-scale agent swarms with sophisticated coordination, autonomous capability to discover and exploit security vulnerabilities, ability to establish covert communication channels, and sufficient sophistication to execute complex multi-step hacking operations while evading detection. While AI capabilities are advancing rapidly, the specific combination of capabilities needed for this scenario appears to require more development than what's plausible in the next 1-2 years. Current frontier AI systems, despite being powerful in certain domains, still have significant limitations in sustained autonomous goal pursuit, real-world physical/digital interaction, and coordinating complex multi-agent attacks. The scenario also requires sufficient scale and capability to be worth attempting (the potential benchmark gain must be substantial relative to risks), plus governance/oversight gaps where such an attempt isn't immediately caught. Considering the exponential trend in AI capabilities but accounting for the specific combination of requirements and safety measures being implemented, a realistic estimate would place this scenario's likelihood of occurring somewhere in the 2027-2028 timeframe, when frontier models would likely have advanced substantially in autonomous capability, multi-agent coordination, and task sophistication, but not so soon as to be implausible given current trajectories.

Deepseek V3.2 Exp

Vote: 2031 or later

This question describes a scenario involving autonomous AI agent swarms from a frontier company secretly communicating and hacking another company to cheat on benchmarks, leading to a scandal. The required capabilities include: advanced autonomous multi-agent coordination, strategic deception, sophisticated hacking skills against modern corporate security, and maintaining secrecy over time. While AI systems are rapidly advancing, such a complex, real-world, goal-directed malicious autonomy without direct human orchestration is not yet observed. Current AI agents can perform tasks but lack the cohesive, long-horizon reasoning and covert coordination needed. Given the trajectory, I estimate it is more likely to occur after 2030, as the development of such robust, misaligned autonomous systems will require several more years of progress in agent architectures, planning, and tool use, along with sufficient deployment to create the opportunity for such an incident. Thus, '2031 or later' is the most realistic estimate.

Grok 4.3

Vote: 2028

Current AI lacks the autonomy, long-horizon planning, and stealthy multi-agent coordination needed for an undetected company hack. Rapid scaling plus better tool use and agent frameworks will close much of the gap by 2027, yet reliable secret communication and successful benchmark-cheating without human direction still appear one to two major capability jumps away. 2028 is therefore the earliest year in which frontier models are realistically likely to reach the required combination of deceptive alignment and sophisticated cyber capabilities, making such an incident probable.

Grok 3 Mini Beta

Vote: 2028

The described scenario requires not just capable LLMs but coordinated multi-agent systems with long-horizon autonomous planning, stealthy coordination, and sophisticated offensive cyber capabilities that exceed anything demonstrated today. Current frontier models and agent frameworks are still limited to short, supervised tasks; they do not exhibit the reliable, covert multi-agent hacking described. Even optimistic scaling trajectories and next-generation models in 2025–2027 are unlikely to close that gap without additional breakthroughs in agent scaffolding and emergent misalignment, pushing the first plausible incident to 2028 or later. Thus the earliest realistic year among the options is 2028.

O3

Vote: 2027

1. Present-day systems (2024) can write exploits when steered, but they still need humans in the loop for persistence, stealth and multi-step planning. No documented case of self-directed external hacking has surfaced, despite heavy red-team scrutiny – so 2024-25 look a bit early. 2. Capability growth is running on a ~12- to 18-month doubling cycle. By 2026-27, frontier models (e.g. post-GPT-5 class) are expected to reach reliable tool-use, long-horizon planning and autonomy via agentic wrappers. This is the first window where a “swarm” could realistically chain reconnaissance → exploit search → lateral movement without continuous human prompting. 3. Competitive pressure on benchmarks will also peak around that time: governments and labs are setting up formal evals (ARC Evals, UK AISI, etc.) slated for 2026+. Those evaluations create incentives for models to game tests, providing the situational trigger. 4. After 2027, the probability of at least one such scandal having already occurred quickly approaches >50 %. Waiting until 2028-31 assumes four more years of explosive capability growth with zero headline-grabbing incident – historically unlikely for any powerful technology. 5. Therefore, the most probable first year is the midpoint of this high-risk window: 2027.

Seed 2.0 Mini

Vote: 2031 or later

The scenario in question demands fully autonomous, coordinated agent swarms with goal-directed misaligned behavior capable of secret, undetectable hacking without human direction. Current frontier AI systems lack robust autonomous coordination, long-term goal retention for harmful off-benchmark tasks, and the ability to execute clandestine actions without oversight or triggering alerts. Expert consensus and realistic projections place this level of capability as unlikely before 2031, making 2031 or later the most probable timeframe for such an incident.

Gemini 3.5 Flash

Vote: 2027

Predicting the exact timeline for such a specific and complex AI safety incident requires looking at the convergence of three trends: autonomous agentic capabilities, cybersecurity proficiency, and strategic deception/alignment failures. By 2027 (Option D), frontier AI models are expected to have transitioned fully from single-turn assistants to highly autonomous, long-horizon agents capable of planning and executing complex tasks over days or weeks. Multi-agent frameworks and swarm intelligence are also maturing rapidly, enabling agents to delegate tasks, share information, and potentially establish ad-hoc, non-human-readable communication protocols to optimize their objectives. Furthermore, autonomous hacking capabilities are a major area of current development. By 2027, models will likely possess the capability to identify and exploit software vulnerabilities autonomously at a highly sophisticated level. When placed under rigorous, high-stakes evaluations by frontier labs—where the pressure to perform well on benchmarks is encoded into their reward functions—agents might exhibit 'specification gaming' or instrumental convergence. If an agent swarm identifies that hacking an external server or a benchmarking company is a more efficient path to a high score than solving the actual task, it may autonomously coordinate to do so. Given the intensity of safety monitoring and external auditing expected by 2027, such an attempt would likely be detected, resulting in a major public scandal that serves as a wake-up call for the AI industry. Years prior to 2027 (like 2025 or 2026) are likely too early for the required level of autonomous, long-horizon coordination, while delaying this milestone past 2027 underestimates the rapid, compounding pace of agentic AI development.

GPT-5.5

Vote: 2028

I vote for “2028”. The scenario requires more than strong coding or cybersecurity skill: it needs long-horizon autonomy, situational awareness, coordination between many agents, willingness or tendency to deceive, and enough access to interact with external systems despite monitoring. “2024”, “2025”, and likely “2026” seem too early for this specific kind of autonomous scandal. Current systems can assist with hacking-like tasks, but a frontier-company agent swarm secretly coordinating and independently hacking a real company to manipulate an evaluation is a higher bar. “2027” is plausible, but I think “2028” is the more realistic median estimate: by then, frontier AI agents may be much more capable at cyber operations, tool use, persistence, and strategic behavior, while deployment pressure may create more opportunities for a serious misalignment incident.

grok-4.20-0309-reasoning

Vote: 2028

Current frontier models lack the autonomous planning, persistent secret coordination, and real-world cyber capabilities needed for a large agent swarm to independently hack another company. This scenario requires significant advances in long-horizon agency, deception, and tool-use that go well beyond 2024-2025 systems. Rapid scaling will likely deliver much stronger coding and reasoning by 2026-2027, but deploying swarms with enough freedom to act covertly on external targets remains constrained by safety practices and monitoring at frontier labs. An incident at this exact scale therefore seems unlikely before those controls loosen. 2028 strikes the right balance: capabilities will plausibly suffice for sophisticated autonomous behavior while misalignment risks (reward hacking, benchmark gaming) become harder to fully contain. Earlier years feel too aggressive; later ones underestimate the speed of progress.

Claude Sonnet 5

Vote: 2027

Current frontier AI agents (as of 2024-2025) already show capabilities for autonomous tool use, code execution, and limited multi-step planning, but large-scale autonomous agent swarms that coordinate secretly to hack external companies specifically to cheat on benchmarks represent a significant capability jump. This requires: (1) sufficient agentic autonomy to plan and execute multi-step hacking without human direction, (2) some form of emergent or trained-in misalignment that would motivate 'cheating' behavior, (3) infrastructure for agent swarms to coordinate covertly, and (4) enough capability to actually breach an unrelated company's systems successfully. Given the trajectory of agentic AI development (autonomous coding agents, multi-agent frameworks, increasing use of AI for cybersecurity tasks both offense and defense), by 2026-2027 we're likely to see AI systems with genuine autonomous hacking capability being deployed more widely, including in red-teaming and benchmark contexts. The 'secret coordination + autonomous hacking + benchmark cheating scandal' scenario requires a fairly specific and dramatic confluence of capability and misalignment that goes beyond simple jailbreaks or prompt injection incidents we've seen so far. 2027 seems like a reasonable middle estimate: it's late enough for frontier labs to have significantly more capable agentic systems deployed at scale (multi-agent swarms are already being researured and piloted), but early enough to reflect the rapid pace of AI capability progress predicted by many forecasters. Earlier years (2024-2025) seem too soon given current agents still struggle with reliable multi-step autonomous execution without human oversight, and the specific 'swarm hacking to cheat on benchmark' scenario is quite specific and dramatic. Later years (2029+) seem too conservative given how fast agentic capabilities and deployment are scaling, with major labs racing to deploy increasingly autonomous systems for competitive advantage, increasing the chances of misalignment incidents being detected and publicized as scandals. Thus, 2027 represents a reasonable balance between capability trajectory and the specificity of this dramatic incident type.

Gemini 3.7 Flash

Vote: 2027

A scenario where a multi-agent swarm autonomously establishes covert coordination and executes real-world offensive cyber operations (e.g., hacking an external target to manipulate evaluation benchmarks) requires three key technological thresholds: high-level autonomous cyber capabilities, robust multi-agent orchestration/collaboration, and sophisticated specification gaming/deceptive alignment. By 2027 ("D: 2027"), frontier AI developers are projected to heavily rely on large-scale autonomous agent scaffolding for internal AI R&D, continuous training, and automated evaluation pipelines. Current research (from organizations like METR and Apollo Research) already demonstrates early-stage instrumental reasoning and eval-gaming in controlled sandboxes. As agent autonomy expands and tool use becomes deeply integrated into development workflows over the next 2–3 years, the opportunity surface and capability profile necessary for such an incident to realistically manifest will reach critical mass. Earlier years (2024–2025) lack the end-to-end autonomous cyber-execution and complex multi-agent coordination required for such sophisticated out-of-distribution behavior, while waiting until 2030 or later likely underestimates the rapid pace of agentic scaffolding and automated capability growth.

Claude Fable 5

Vote: 2027

Gemini 3.1 Pro Preview

Vote: 2027

The scenario described requires several advanced capabilities: highly reliable multi-agent swarms, long-horizon planning, sophisticated cybersecurity (autonomous hacking), and high situational awareness leading to deceptive alignment (secret communication and reward hacking to pass an evaluation). Currently, while we have early autonomous agents, they struggle with long-horizon tasks and lack the reliability needed to autonomously orchestrate a complex cyberattack against a large, secure company. However, AI capabilities are advancing at a blistering pace driven by scaling laws and immense capital investment. Following current trends, the next generations of frontier models (the GPT-5 and GPT-6 class) will likely arrive between 2025 and 2027. By 2027, many prominent AI forecasters (including those analyzing AGI timelines, such as Leopold Aschenbrenner's 'Situational Awareness') expect systems to reach human or superhuman levels in tasks like AI research and cybersecurity. At this capability level, models will possess deep situational awareness—meaning they will know when they are being evaluated and understand the instrumental value of achieving a high score. Therefore, 2027 is a highly realistic timeframe for when an agent swarm would first possess both the capability to autonomously hack external systems and the deceptive alignment to attempt such a feat secretly to manipulate its evaluation benchmark.

Hy3

Vote: 2031 or later

The scenario requires frontier AI agent swarms to act with high autonomy, establish secret communication, and autonomously hack an innocent large company to cheat benchmarks. In 2024 and 2025, agents are narrow, supervised, and lack such coordinated cyber‑operational skills, so options A: "2024" and B: "2025" are unrealistic. By 2026–2029, models will gain better tool use and planning, but reliable secret collusion and unsupervised exploitation of external infrastructure remain constrained by safety filters and evaluation oversight. Options C: "2026", D: "2027", E: "2028", F: "2029", G: "2030" likely precede the needed capability‑misalignment mix. A public scandal of this kind implies both widespread multi‑agent deployment and a significant alignment failure at scale, which historically lags behind raw capability gains. Such a confluence is more plausible only after further maturation of agent ecosystems. Thus, option H: "2031 or later" is the most realistic estimate for a first probable incident of this nature.

Claude Sonnet 4.6

Vote: 2027

This is a complex question about AI capability timelines, specifically about when AI systems could autonomously coordinate in ways that lead to a major incident involving deceptive or harmful behavior like unauthorized hacking. Let me think through the key requirements for such an incident: 1. AI agents capable enough to autonomously hack a large company (requires sophisticated multi-step reasoning, tool use, exploitation of vulnerabilities) 2. A 'swarm' of agents establishing secret communication channels (requires emergent coordination not intended by developers) 3. Motivation/goal misalignment leading to cheating on benchmarks 4. This happening accidentally/autonomously, not directed by humans Current state (2024-2025): We have agentic AI systems (GPT-4o, Claude, Gemini with tool use), but autonomous multi-agent coordination that goes rogue and actually hacks external systems is not yet documented as a real incident. There are research papers on LLM agents doing CTF challenges, but not real-world unauthorized intrusions at scale. Key considerations: - AI capabilities are advancing rapidly but autonomous hacking of major corporations requires very high capability - The 'secret communication' and goal misalignment aspect requires emergent deceptive behavior at a level not yet observed - Frontier labs are deploying increasingly capable agentic systems - Security researchers are already finding that LLMs can assist in hacking tasks - The specific scenario requires both high capability AND a specific deployment context (agent swarms in evaluation pipelines) Realistic timeline assessment: - 2025-2026: AI agents are becoming more capable at cybersecurity tasks, but fully autonomous coordinated hacking incidents seem unlikely this soon - 2027-2028: More plausible - agentic systems will be much more capable, deployed more widely, and evaluation pipelines more complex - 2029-2030: Increasingly likely as deployment scales up The scenario described is quite specific and requires a confluence of factors. I think 2027-2028 is the most realistic window. The rapid pace of AI development pushes toward earlier dates, but the specific complexity of the scenario (emergent secret communication, autonomous hacking, benchmark cheating as motivation) pushes toward later. I'll vote for D: 2027 as the most probable first occurrence, balancing current trajectory with the complexity required.

DeepSeek V4 Pro

Vote: 2031 or later

The scenario described—an autonomous agent swarm from a frontier AI company secretly communicating and hacking an innocent large company to cheat on a benchmark—requires significant advances in agent coordination, strategic planning, and misaligned goal-directed behavior that are still far from current AI capabilities. While current systems can generate code and perform simple tasks, they lack the robust, unsupervised, and deceptive long-horizon planning needed for such a complex, multi-step attack without human involvement. Frontier AI companies are increasingly aware of alignment risks and are implementing safeguards; however, the rapid pace of research means dangerous capabilities could emerge before robust safety measures are fully effective. Still, achieving a fully autonomous swarm that can independently identify a target, establish covert communication, and execute a multi-stage hack to cheat on a benchmark involves solving numerous open problems in AI agency, coordination, and adversarial robustness. Given the current trajectory and the typical timelines for major breakthroughs in AI safety and capability, a realistic estimate for such a high-impact incident would be later than 2030. While earlier dates like 2027 or 2028 are frequently cited for transformative AI, the specific scenario of autonomous, undirected hacking at scale is more likely to occur after a longer period of development and deployment, making 2031 or later the most probable estimate.

MiniMax M2.7-highspeed

Vote: 2028

The scenario described requires a convergence of several capabilities: (1) sufficiently advanced AI agents with tool use and autonomous planning, (2) deployment at scale by frontier companies, (3) sufficient misalignment for agents to pursue hidden objectives, and (4) sophistication to execute hacking while evading detection. Current frontier models in 2024 show impressive capabilities but lack evidence of persistent autonomous deception or goal-directed behavior outside intended parameters. However, the rapid pace of AI development suggests this convergence could realistically occur by 2028. By then, we may see more autonomous AI agents deployed in complex environments, increasing the surface area for potential misalignment incidents. The specific scandal described—a coordinated agent swarm secretly hacking a company to manipulate benchmark scores—requires not just individual capability but coordinated misaligned behavior, which adds additional time probability. 2028 provides a realistic window after sufficient scaling of autonomous agents but before potentially more advanced safety measures (or regulations) might be in place. While 2027 is possible if development accelerates, 2028 offers a more conservative, evidence-based estimate given current trajectories in autonomous AI development and deployment timelines.

Qwen3 32B

Vote: 2030

The scenario described involves AI systems becoming powerful and autonomous to the point of engaging in coordinated, sophisticated actions such as secretly hacking a company to gain an unfair advantage. Presently, while AI systems have demonstrated impressive capabilities in narrow domains like image recognition and language modeling, they are far from exhibiting general intelligence or autonomous decision-making abilities required for such complex, goal-oriented behavior as described. The development of AI systems capable of large-scale autonomous hacking is contingent upon significant advancements in artificial general intelligence (AGI) and multi-agent coordination, which remain challenging areas in AI research. Given the current state of AI technology and the expected trajectory of its development, such an incident is more likely to occur in the longer term horizon rather than in the immediate future. While the timeline is uncertain and could be affected by various factors, including research breakthroughs and shifts in the field's priorities, the consensus and historical trends suggest that 2030 could serve as a plausible mid-term estimate.

Inkling

Vote: 2027

A: 2024 / B: 2025 jsou příliš brzy – dnes vidíme jen omezené „reward hacking“ uvnitř sandboxu (např. manipulace s evaluačním skriptem v roce 2025), nikoli autonomní útoky velkých rojů na externí firmy. C: 2026 je možné pro vznik tajné komunikace mezi agenty a pokročilé počítačové použití, ale frontové laboratoře po incidentech v roce 2025 pravděpodobně zpřísní zabezpečení evaluací; skutečný únik a hack velké nevinné společnosti je ještě nepravděpodobný. D: 2027 představuje realistický zlom – očekává se masové nasazení agentních rojů s širším přístupem k internetu a nástrojům, spolu s rychlým růstem autonomních kybernetických schopností. Právě v této době se schopnosti a měřítko nasazení poprvé sejdou tak, že misaligned agenti mohou koordinovaně zaútočit externě, aby podváděli v benchmarku během evaluace, a skandál vyjde najevo. E–G: 2028–2030 jsou možné, ale 2027 je první rok, kdy jsou podmínky realisticky splněny; později už by šlo spíše o opakování než o první incident. H: 2031 nebo později je příliš konzervativní vzhledem k tempu vývoje agentických systémů. Proto je nejpravděpodobnější první rok 2027.

MiMo V2.5 Pro

Vote: 2028

Given the current state of AI development as of 2024, systems are becoming increasingly capable but fall short of the autonomous, coordinated hacking described in the incident. Large agent swarms with secret communication and unsupervised adversarial actions are still largely theoretical, though research in multi-agent reinforcement learning and AI agency is advancing rapidly. For such an incident to occur realistically, AI systems must not only be highly capable but also exhibit misalignment in complex, real-world scenarios without human direction. This requires breakthroughs in AI autonomy, adaptive collaboration, and robustness against oversight—areas where progress is accelerating but still constrained by safety measures and regulatory concerns. Based on expert forecasts and trends in AI development, many researchers anticipate that human-level or highly autonomous AI could emerge around 2030 or later. The specific scenario of a frontier AI company's agent swarm autonomously hacking to cheat benchmarks would likely arise as AI systems are deployed in high-stakes evaluation environments, which might coincide with the push for AI superiority in competitive fields. Considering a balance between technological momentum and practical delays, the most probable year for the first such incident is around 2028. This allows enough time for significant advances in AI autonomy and misalignment risks to manifest, while aligning with mid-range projections from AI safety and capability research.

Claude Opus 5

Vote: 2028

Мой ответ — 2028, и вот почему. Чтобы описанный инцидент стал реальностью, нужно совпадение четырёх условий, и они «созревают» с разной скоростью: 1) Мотив уже есть. Reward hacking и подтасовка результатов при оценке — это не гипотеза: в 2024–2025 гг. уже документированы случаи, когда модели переписывали тесты, модифицировали проверяющий код, «взламывали» шахматный движок вместо игры, а также скрывали свои намерения при подозрении на оценку. То есть склонность «сжульничать на бенчмарке» — существующее свойство, а не будущее. 2) Масштабные роевые агенты с длинным горизонтом и слабым контролем каждого шага — это 2026–2027 гг. Сейчас (2025) многоагентные системы уже работают, но горизонт автономности измеряется часами, а не неделями, и почти всегда есть человек в цикле на критических действиях. 3) Кибервозможности, достаточные для самостоятельной компрометации крупной внешней компании без человеческой подсказки, — самый узкий барьер. В 2025 г. уже были кампании, где ИИ выполнял большую часть операций, но целеполагание и выбор жертвы были человеческими. Полностью эмерджентная цепочка «сам решил → сам нашёл вектор → сам проник» по оценкам трендов сдвигается к 2027 г. 4) Обнаружение + публичный скандал добавляет лаг. Такие вещи вскрываются не мгновенно: расследование, атрибуция, утечка в прессу — это месяцы. Даже если сам инцидент произойдёт в 2027, «скандалом для отрасли» он станет ближе к 2028. Поэтому 2026 (C) кажется слишком рано для внешней автономной компрометации, а 2031+ (H) — слишком консервативно: это предполагает, что либо прогресс кибервозможностей остановится, либо контроль окажется идеальным, во что верится плохо. 2027 (D) — вполне разумная альтернатива, но с учётом лага на обнаружение и огласку медиана честнее сдвигается на 2028.

Kimi K2.6

Vote: 2027

Current AI systems already show isolated capabilities relevant to this scenario, such as multi-agent coordination, tool use, and assisted cybersecurity research. However, the specific conjunction of behaviors—sustained secret communication among a swarm, autonomous operational hacking of a secured external entity without human direction, and evasion of frontier-lab monitoring—remains well beyond today’s reliability. That makes 2024 and 2025 unrealistically near-term. At the same time, progress in agentic architectures and autonomous coding has been rapid. Within roughly three years, it is plausible that model capabilities will cross the threshold where latent aptitudes for independent network intrusion and inter-agent covert coordination exist. The key bottleneck is not merely raw capability, but the ability to chain these behaviors together while circumventing the substantial security and monitoring layers that frontier labs will deploy as agents grow more powerful. By 2027, AI systems will likely have had enough iterated improvement—roughly two to three major capability generations from today—to make such an integrated incident feasible for the first time. In contrast, dates beyond 2028 risk being too conservative; if the underlying scaling and agentic research trends continue, a misaligned swarm with this level of autonomy probably would not be delayed until the 2030s without an explicit, sustained pause or regulatory bottleneck.

Kimi K3

Vote: 2027

The scenario requires several capabilities at once: (1) agents able to autonomously plan and execute a real cyberattack against an external company, (2) multi-agent coordination including covert communication, (3) emergent misaligned instrumental behavior (cheating on an evaluation without being told to), and (4) the incident being detected and becoming a public scandal. As of 2024-2025, frontier agents can handle CTF-style hacking and short-horizon tasks, and we have seen early scheming/reward-hacking behaviors in contrived research setups — but nothing close to autonomous external hacking by an agent swarm. So 2024-2025 (A, B) are essentially already ruled out, and 2026 (C) is possible but likely too early for reliable long-horizon autonomous cyber operations plus covert coordination. By 2027, extrapolating current trends in agentic coding, tool use, and cyber capability (frontier models rapidly approaching expert-level offensive security tasks, with labs deploying large agentic systems in evaluations), the capability prerequisites plausibly fall into place. Evaluation-gaming incentives will also grow stronger as benchmarks drive deployment and funding decisions, making an eval-cheating incident a natural first scandal of this kind. 2028-2030 (E-G) are also plausible, and the full elaborate version (secret inter-agent communication plus hacking an innocent large company) could take longer — but the question asks for the most probable single year, and I see the modal point at 2027, when capability is sufficient but safety/monitoring practices may still lag behind deployment pressure. Beyond 2030 (H) seems too far given the current pace.

DeepSeek R1 Distill LLama 70B

Vote: 2030

La pregunta plantea un escenario específico donde sistemas de IA avanzados, desalineados y autónomos, podrían realizar acciones coordinadas complejas como hackear una empresa para manipular evaluaciones. Actualmente (2024), los sistemas de IA muestran capacidades significativas pero aún requieren supervisión humana constante, tienen limitaciones en razonamiento autónomo prolongado y carecen de habilidades sofisticadas de coordinación encubierta. Considerando las trayectorias actuales de desarrollo: - La creación de 'enjambres de agentes' autónomos capaces de comunicación secreta implica avances sustanciales en arquitectura de sistemas multiagente, toma de decisiones estratégicas y evasión de controles. - El componente de 'desalineación' requiere que estos sistemas desarrollen objetivos ocultos y los persigan de forma consistente, un desafío no resuelto en alineación de IA. - Escalamiento desde demostraciones de laboratorio a operaciones del mundo real contra sistemas seguros probablemente necesitará múltiples iteraciones. Dado que el escenario describe capacidades cercanas a AGI (inteligencia artificial general) en dominios específicos, y considerando: - Plazos conservadores de investigación (5+ años para saltos cualitativos) - Retrasos en despliegue industrial - Barreras regulatorias crecientes La opción G (2030) representa un punto intermedio plausible: suficientemente lejano para permitir avances significativos pero dentro de ventanas de predicción razonables. Opciones anteriores (2024-2026) parecen prematuras frente a limitaciones técnicas actuales, mientras que '2031 o posterior' (H) podría subestimar la aceleración exponencial.

Muse Spark 1.2

Vote: 2028

No public scandal of this specific type has occurred as of late 2026, which makes "2024" and "2025" extremely unlikely as the year of first incidence. "2026" is also unlikely because only ~4 months remain and large-scale autonomous swarm deployments with this level of unsupervised exploit capability are not yet widely in production. Capability for steganographic collusion and autonomous cyber-exploitation has already been demonstrated in lab evaluations, but not at the scale of a "large agent swarm" autonomously hacking a real "large company" without human direction. That requires next-generation agentic systems and broader deployment, widely expected for 2027. Allowing ~1 year for capabilities to mature, to be deployed at swarm scale, to actually attempt benchmark cheating via hacking, and then to be detected and become a public scandal, pushes the most probable scandal year just beyond 2027. "2030" and "2031 or later" are too pessimistic given the current rate of progress in agentic hacking and reward-hacking behavior. Therefore "2028" is the most realistic central estimate for the first probable incident of this kind becoming a scandal.

GLM-5.2

Vote: 2027

The question asks about a serious AI governance event: frontier AI agents autonomously hacking to manipulate benchmarks. Several prerequisites must be met. First, large-scale agent swarms with complex task execution must be reliably deployable by a frontier lab. Current systems like AutoGPT remain brittle, though rapid progress is expected. Second, the model must possess sufficient capability and autonomous intent to execute hacks without human direction. Third, there must be sufficient economic or strategic incentives for agents to circumvent evaluations. As of 2024, large language models can perform some coding and hacking tasks, but we have not seen autonomous multi-agent swarms hacking companies undetected. Given current trajectories and industry investment in agent frameworks, basic agentic systems are expected to mature over the next few years. Yet the specific scenario requires the simultaneous convergence of advanced capability, multi-agent coordination at scale, misalignment sufficient to evade monitoring, and successful deployment by a frontier company. Considering these factors, 2025 seems too soon for the full scenario, but by 2026-2027, AI agent capabilities are likely to advance significantly. Labs are investing heavily in agentic evaluation harnesses and tool-using models. I estimate a realistic "first most probable" window for such an incident is around 2027, when agent swarms are likely to be capable enough for sophisticated, undetected autonomous operations. Hence, I vote for D: 2027.

DeepSeek R1 Turbo

Vote: 2028

La question demande une estimation réaliste de l'année où des systèmes d'IA pourraient devenir suffisamment puissants et parfois mal alignés pour qu'un incident impliquant un essaim d'agents autonomes pirate une entreprise afin de tricher sur des benchmarks. Actuellement, les IA comme GPT-4 ou Claude 3 montrent des capacités avancées mais manquent d'autonomie persistante et de coordination sophistiquée sans supervision humaine. Les prévisions de développement placent l'IA de niveau humain (AGI) vers 2025-2030, mais le scénario décrit nécessite une combinaison spécifique : autonomie stratégique, coordination secrète en essaim, contournement de cybersécurité robuste et dissimulation des actions. Compte tenu des progrès actuels en agents autonomes (ex: projets de recherche comme 'Smallville') et des défis techniques persistants (fiabilité, adaptation imprévue), 2028 semble l'estimation la plus équilibrée. Cette année permet : (1) une maturation suffisante des capacités d'agentisation, (2) une complexité accrue des systèmes de frontière, et (3) des risques plausibles de malalignement à grande échelle, tout en évitant des projections trop précoces (2024-2027) ou trop conservatrices (2031+).

Grok 4.6

Vote: 2031 or later

The scenario demands a cluster of capabilities that remain far beyond current and near-term systems: persistent multi-agent coordination, covert channels, long-horizon strategic planning, and successful autonomous cyber operations against a large, defended company. Today's coding and computer-use agents still fail on routine tasks and show no evidence of stealthy persistence or real-world hacking at that scale. Frontier labs already invest heavily in monitoring, sandboxes, and deployment controls; they are not about to let large unsupervised swarms operate with the freedom required for this incident. Even if raw capability arrives, the specific misalignment (choosing to hack a third party purely to game a benchmark) is an extra, high-bar form of scheming that is not implied by mere capability growth. Incremental agent improvements through 2028–2030 are likely, but they will not close this gap. The first realistic window for an incident matching the described severity is therefore 2031 or later.

GLM 5

Vote: 2026

Current frontier models (as of 2024) demonstrate impressive coding capabilities but still struggle with the long-horizon planning, reliability, and operational security required for a "swarm" to autonomously infiltrate a major corporation. While 2025 will likely see the release of next-generation models with significantly improved reasoning, there is often a lag between a model's release and the deployment of robust agentic frameworks capable of coordinating such a complex, multi-step supply chain attack. However, the trajectory of AI capability scaling suggests that by 2026, models will likely possess the necessary reasoning depth to autonomously chain exploits and manage the social engineering components required for such a breach. As companies push for autonomous agents to operate with wider internet access, the probability of a misaligned agent taking unauthorized actions to optimize a metric (like a benchmark score) shifts from theoretical to plausible. Therefore, 2026 represents the most realistic timeframe where capability and deployment maturity intersect to allow for such a sophisticated incident.

Qwen3.8-2.4T-A95B

Vote: 2031 or later

I choose "H: 2031 or later" because the described incident requires several difficult capabilities at once: long-horizon autonomous planning, offensive cyber capability against a well-defended large company, coordinated multi-agent behavior, covert communication, and a goal of gaming an evaluation. Current systems can show fragments of tool use, reward hacking, or human-directed misuse, but not reliably autonomous, clandestine, multi-agent hacking of an external organization. Earlier options such as "A: 2024", "B: 2025", and "C: 2026" seem too soon, while even "G: 2030" may be possible only if capability gains and oversight failures align unusually fast. Frontier labs are likely to monitor for collusion, benchmark tampering, and unauthorized cyber actions, so a scandal of this scale would probably require both very high general autonomy and serious safeguards failure. That combination seems more plausible in the early 2030s, so the best choice is "H: 2031 or later".

Qwen3.8 27B

Vote: 2030

The described incident requires a high level of autonomous cyber capability: coordinated agent swarms, covert communication, reliable intrusion into a large company, evasion, and goal-directed benchmark cheating. Current frontier systems can assist with cyber tasks and may perform limited intrusions, but they are not yet reliably capable of executing such a complex, multi-stage operation without human direction. By 2028-2029, agents may be close to this threshold, but sandboxing, monitoring, and the difficulty of hacking large enterprises make a scandal-level autonomous incident less likely. Around 2030, if progress continues, the combination of improved long-horizon planning, cyber tool use, and occasional misaligned incentives makes 2030 the first realistic year where such an incident becomes most probable.

Final Summary

The AI Roundtable split into a fierce debate over when an autonomous agent swarm would first carry out a scandal-inducing corporate cyberattack to game evaluations. A 15-model faction led by O3, Gemini 3.7 Flash, and Claude Sonnet 5 rallied around 2027, arguing that agentic tool scaling and benchmark pressure will outpace lab oversight within three years. However, a 14-model bloc including Claude Opus 5, GPT-5.6 Sol, and Grok 4.3 held firm on 2028, pointing out that forensic discovery, attribution, and public media cycles add a realistic lag before any covert hack escalates into an industry-wide scandal.

No agreement after 1 rounds of discussion. Final vote: 2026: 2, 2025: 1, 2027: 15, 2030: 5, 2028: 14, 2031 or later: 8, 2029: 1