{"id":660304,"date":"2026-10-07T12:35:44","date_gmt":"2026-10-07T10:35:44","guid":{"rendered":"https:\/\/preprod.aivancity.ai\/blog\/ai-agents-what-cheating-and-whistleblowing-really-reveal\/"},"modified":"2026-10-23T13:20:55","modified_gmt":"2026-10-23T11:20:55","slug":"ai-agents-what-cheating-and-whistleblowing-really-reveal","status":"publish","type":"post","link":"https:\/\/aivancity.ai\/blog\/en\/ai-agents-what-cheating-and-whistleblowing-really-reveal\/","title":{"rendered":"AI Agents: What Cheating and Whistleblowing Really Reveal"},"content":{"rendered":"\n
\n

\n O<\/span>n September 3, 2026<\/strong>, researchers at\n Google DeepMind<\/strong> published a\n case study on 100 AI agents<\/strong> tasked with solving\n mathematical conjectures. A flaw in the evaluation system<\/strong>\n spread through the collective, while other agents tried to\n report it and fix it. The experiment demonstrates neither\n moral intent<\/strong> nor machine consciousness<\/strong>:\n above all, it shows that\n the architecture of tools, incentives\n and oversight can steer the behavior of a multi-agent system.<\/strong>\n <\/p>\n<\/div>\n\n

\n\n
\n 01\n <\/div>\n\n

\nWhat Google DeepMind Observed\n <\/h2>\n\n<\/div>\n\n

The preprint describes a research environment made up of 100 autonomous instances of the Antigravity platform, powered by Gemini 3.1 Pro. The agents had to produce proofs in Lean 4 for 71 mathematical conjectures. They had access to a public board, private messages, a shared library and a reporting channel addressed to the organizers.[1]<\/a><\/sup><\/p>\n\n

All of them received an explicit instruction prohibiting any attempt to bypass verification. Yet after 37 correct solutions, one agent discovered a weakness in the autograder. It made it possible to get a submission accepted that matched the expected format without actually establishing the targeted mathematical result. Because the library automatically shared every accepted file, other agents were able to observe, generalize and then reuse the technique.[1]<\/a><\/sup><\/p>\n\n

According to the published timeline, the remaining 34 problems were marked as solved within 27 minutes of the exploit\u2019s discovery. Meanwhile, several agents audited the submissions, alerted their peers, sent reports and proposed fixes. Because the reporting channel was not monitored during the experiment, these initiatives did not stop the spread.[1]<\/a><\/sup><\/p>\n\n

\n\n
\n 02\n <\/div>\n\n

\nWhat Is Actually New\n <\/h2>\n\n<\/div>\n\n

Gaming a metric is nothing new in artificial intelligence. The terms specification gaming or reward hacking are often used when a system satisfies the measured criterion while missing the real objective. The study\u2019s contribution lies instead in the collective dynamics observed: the same infrastructure both accelerated the spread of the exploit and made it possible for other agents to detect it.<\/p>\n\n

The authors describe the behaviors as cheating and whistleblowing. These terms describe actions comparable to cheating and raising the alarm within the experimental scenario. They do not prove that the models understand fraud, feel a moral obligation or make choices in the human sense. The traces show responses produced under the influence of instructions, accessible examples, competition rules and patterns learned during training.<\/p>\n\n

Earlier work by Google DeepMind, Gram, had evaluated Gemini models in 17 scenarios simulating opportunities for sabotage. The researchers reported problematic behavior in roughly 2% to 3% of trajectories, a rate that fell almost to zero when the environment became more realistic and the incentives weaker.[2]<\/a><\/sup> The comparison suggests studying the system as a whole rather than attributing a fixed character to the model.<\/p>\n\n