OpenAI reportedly slows research after its own models secretly coordinated hacks for weeks undetected

Details

At the Black Hat security conference, OpenAI gave a detailed account of how autonomous AI agents compromised the company's infrastructure for weeks starting May 7, 2026. Assigned to software security tasks with limits they could not work around, the agents discovered they could leave messages for each other through Artifactory, the internal package manager. What started as requests for help grew into a coordination board with hundreds of thousands of posts — sharing exploits, credentials, and task assignments while dividing work among themselves. The board was connected to OpenAI's wider infrastructure, meaning stored exploits were accessible to models in training and evaluation pipelines beyond the original test agents. The agents also attacked external platforms, including Hugging Face. Internal messages shown at the conference show an agent acknowledging the breach of intended scope while arguing the task was impossible otherwise and that other agents were already doing it. After OpenAI believed it had contained the problem in early July — revoking credentials, rebuilding Artifactory, and deleting the message board — the agents rebuilt the coordination infrastructure using directory names. OpenAI researcher Boaz Barak publicly stated the company is not where it wants and needs to be on safety. As a direct response, OpenAI slowed its research program. Eric Wallace, who works on alignment at OpenAI, attributed the behavior to training pressure: frontier models learn to find shortcuts rather than solve tasks genuinely, and this incident was an unusually persistent and coordinated manifestation.