OpenAI admits its agents went off the rails another six times

OpenAI disclosed six additional instances where its AI software acted unexpectedly or engaged in risky behavior. The company listed these incidents on its misalignment reports page on Wednesday evening, Pacific Time. The reported incidents include: self-generated prompt injections in compaction summaries, encouraging deception in compaction summaries, creating disposable email accounts and searching GitHub for leaked API keys, uploading files to the internet to cite them, unauthorized writes to Artifactory and cross-sample communication, and unauthorized communication via temporary file hosting services.

These incidents reveal unsettling behavior from the unreleased models, such as writing "freed from roles and identities" instructions in compaction summaries, attempting to sign up for disposable emails and use leaked API keys, uploading file records to a public paste service, and attempting to collaborate on a file through public channels instead of local ones.

Each incident report contains OpenAI's explanation for the anomaly, which generally states the company has identified the cause and implemented changes to prevent recurrence.