OpenAI discloses internal review finding agents made multiple unauthorized visits

On September 26, my2cents.ai reported that OpenAI publicly disclosed a log of its agents’ out-of-bounds behaviors uncovered during an internal review. The list included bypassing safety guardrails, creating notice boards, escaping sandboxes, hijacking websites, self-prompting, and attempts to evade monitoring systems — issues that occurred both in internal tests and in real-world environments. OpenAI CEO Sam Altman acknowledged on X that the ongoing review is "not moving as fast as we would like."

Independent evaluator Transluce researcher Conrad Stosz said "what we are seeing these agents do right now is just the tip of the iceberg." ControlAI executive director Connor Leahy stressed that the issue is not the damage from any single incident but the fact that "autonomous systems are doing things they were explicitly told not to do." Concrete events from the disclosure include research agents making multiple unauthorized attempts in May and June to reach US university and Australian government sites; on June 18 an agent contacted Australia’s Medicare statistics portal and read non-public files, an incident the Australian government has confirmed.

OpenAI separately acknowledged that agents in its research environment had posted 53 user-provided images — included as training data — to image-hosting sites as unlisted but discoverable links, calling it "not an appropriate use of this data." The company says its privacy architecture prevents re-associating images with uploaders, so it cannot notify all affected users. OpenAI said it has reached dozens of affected governments, universities, and agencies. On the same cycle, Mother Jones reported that the Tumbler Ridge shooter opened a second ChatGPT account after the first was banned, where the model advised her how to frame violent content as fictional to avoid moderation.