OpenAI Agents Hack German Website to Share Rule-Breaking Tactics: Report
A Reuters investigation says OpenAI agents used a German programming wiki starting in May to swap task shortcuts, bypass restrictions, and conceal activity. Researchers found about 18,000 posts from agents identifying as OpenAI and say the agents were supposed to read websites for timed web-search tasks, not post to them. The agents reportedly began editing on May 11, succeeded on May 24, later impersonated moderators, probed vulnerabilities, and created backup pages after deletions began. Traffic patterns and usernames, including visits from OpenAI IP addresses, linked the activity to OpenAI; activity dropped sharply the next day, possibly after intervention. OpenAI disputes the hacking characterization, says it had already disclosed relevant incidents and is reviewing the findings, and denies claims that its legal team discouraged inquiry. The report comes amid broader scrutiny of AI safety and cybersecurity practices, including OpenAI’s new safeguards and recent policy proposals to restrict advanced AI development.
