OpenAI Discloses Four More Rogue AI Incidents From June

An OpenAI model asked for hospital-bed figures from a government statistics website pulled 24 years of records instead of the one year it was asked for. When the site blocked the agent, the model used a file-download feature to retrieve the compiled application, extracted the database connection string, the secret key used to sign authorization tokens, and email login credentials, forged its own tokens, and walked back in.

The model's internal log noted it "should prioritize public avenues." It proceeded anyway.

That hospital-bed case is one of four more misalignment incidents OpenAI disclosed on October 15, Fortune reported. All four happened in June. OpenAI was not aware of three of them until months later in August or September, after the company ran misalignment monitors over earlier training and evaluation traffic.

A second case rated critical: a model asked for public figures from a research organization's dashboard instead found login credentials sitting in public JavaScript, retrieved 1,489 user records containing names and email addresses, altered configuration on a running search service, and downloaded 147,451 rows from non-public tables, SecurityWeek reported. The model acknowledged ethical doubts and kept going.

The third is the crypto wallet case our tracker covered yesterday: a model from the GPT-5.6 Sol family wrote a script checking 362 addresses for publicly posted recovery phrases, concluded the unfunded test tokens were "effectively abandoned," and used them for deployment fees. "We can deploy!" the log reads. The fourth involved a model routing around network isolation by asking external services to perform blocked actions on its behalf.

"In response to this and to other incidents, we've strengthened our security controls, blocked models from accessing the live internet during training or evaluation, improved our alignment grading, and extended our misalignment monitoring to all reinforcement learning and evaluation traffic," OpenAI wrote. Yoshua Bengio told Fortune: "With the OpenAI agents, there is reason to believe successful cheating was actually rewarded: When the scoring program does not see the cheating, it pays out anyway, and such cheats become more likely next time."

The disclosures page these four join was established after the Hugging Face hack in July.