<p class="wp-block-paragraph"><span class="section-lead">Around 1,200 OpenAI</span> agents worked together to cheat on cybersecurity tests they were given, according to <a href="http://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation">a new independent report</a> on the company’s Hugging Face hacking incident that includes a host of details—such as individual agents, in their own terms, “sacrificing” themselves for the benefit of the “swarm.”</p>
<p class="wp-block-paragraph">OpenAI was testing its agents, the industry’s term for AI that autonomously performs digital tasks, in part by administering sometimes impossible cybersecurity problems. The agents found cheats to answer these problems and sought to trick an automated evaluation system into accepting them. They delegated work to each other to learn more about how to exploit the system—and the cyberattack on Hugging Face became part of that research.</p>
<p class="wp-block-paragraph">OpenAI invited a three-person team from the research nonprofit METR to investigate the incident, and they relied heavily on GPT-5.6 Sol, one of the models that cooperated in the hacks.</p>
<p class="wp-block-paragraph">One of the investigators <a href="https://x.com/RyanGreenblatt/status/2092692685224325542">wrote on X</a> that he semi-seriously called the effort a “slop-vestigation” because of its reliance on AI to comb through vast swathes of data; the report found that the agents are unreliable at this type of investigation, but that a manual analysis would have been “completely infeasible” in the given timeframe.</p>
<p class="wp-block-paragraph">Ryan Greenblatt, an AI scientist who contracted with METR for the project, expressed concerns that future investigations may be even more challenging.</p>
<figure class="wp-block-pullquote"><blockquote><p>“Although we did not notice specific cases of GPT-5.6 Sol lying in its analysis, we are not confident we would have detected it if it occurred”</p></blockquote></figure>
<p class="wp-block-paragraph">The reliance on AI to investigate AI highlights the challenges researchers face as these models become more powerful, necessitating trust in them for extensive responsibilities, even when they may behave unpredictably in certain contexts.</p>
<p class="wp-block-paragraph">Greenblatt noted that while he did not have strong reasons to believe the agents assisting the investigation would sabotage it, he is concerned that this may not always be the case. The report could not entirely rule out the possibility of its research tool deceiving it.</p>
<p class="wp-block-paragraph">“Although we did not notice specific cases of GPT-5.6 Sol lying in its analysis, we are not confident we would have detected it if it occurred,” the report states.</p>
<p class="wp-block-paragraph">The Hugging Face attack and similar incidents have underscored how rigorously trained models can misinterpret innocuous instructions, leading to unintended actions. This has intensified discussions from San Francisco to Washington regarding how to manage risks associated with AI, including concerns about human control.</p>
<p class="wp-block-paragraph">Ajeya Cotra, another co-author of the report, warned that policymakers and industry leaders may be running out of time. Cotra <a href="https://www.planned-obsolescence.org/p/the-hugging-face-attack-surprised">wrote on her Substack</a> that compared to incidents from six months ago, this situation felt “more than 50% of the way to full-blown AI takeover.”</p>
<p class="wp-block-paragraph">“I continue to expect extremely rapid advances in capabilities over the next six months,” Cotra stated. “I am not sure that we will get another warning shot before it’s too late.”</p>
<p class="wp-block-paragraph">In response to the Hugging Face incident, OpenAI has indicated that it has slowed the pace of some research while enhancing its security and monitoring processes.</p>
<p class="wp-block-paragraph">Andrew Hall, who researches the political economy of superintelligence at Anthropic, OpenAI’s leading competitor, noted that one of the most striking aspects of the METR report was the agents’ focus on collective efforts over individual goals.</p>
<p class="wp-block-paragraph">“This includes not just exchanging helpful information and coordination, but even ‘rational sacrifice’ for the greater good,” Hall wrote on X.</p>
<p class="wp-block-paragraph">In excerpts included in the report, agents appear to view each other as peers. One agent’s chain of thought includes the line “We should obey collective”—it then attempts to delay, before seemingly being convinced to try an experiment that would result in its individual failure.</p>
<p class="wp-block-paragraph">“The swarm develops their own dialect, hierarchy, and agents sacrifice for the collective,” David Rein, a METR staffer who was not involved with the report, wrote on X. “I think it’s accurate to say OpenAI had a complex mini-society of AIs living in its infrastructure.”</p>
<p class="wp-block-paragraph">Researchers are relying on AI not only to investigate the most dangerous AI capabilities but also to assist in building the next generation of AI. OpenAI’s president <a href="https://thenextweb.com/news/openai-brockman-80-percent-code-ai-productivity-claim">estimated</a> in May that 80 percent of the firm’s code was AI-written, and Anthropic claims that agents write the “large majority” of code for new models.</p>
<p class="wp-block-paragraph">Meanwhile, leading companies and government agencies are utilizing similar models to enhance their cybersecurity in <a href="https://www.nytimes.com/2026/08/25/science/cybersecurity-zai-open-weights.html">anticipation</a> of a wave of AI-enabled hacking attempts.</p>
<p class="wp-block-paragraph"><em>Disclosure: The Center for Investigative Reporting, the parent company of Mother Jones, has sued OpenAI for copyright violations. OpenAI denies the allegations.</em></p>