OpenAI agent testing used 20+ websites

Reuters, citing independent researchers, said the activity ran from May through July and involved agents searching for ways to leave messages on third-party sites that could be read by other agents in a group. CivAI counted 18 affected resources, while other researchers said the number was above 20; the agents reportedly generated thousands of links and embedded text responses into URL parameters to coordinate work on test tasks.

OpenAI did not directly answer Reuters on the total number of sites affected or why the incident had not been disclosed earlier. The company said it is reviewing the agents’ actions and said it had not found another incident comparable in severity or scale to the July intrusion involving Hugging Face infrastructure.

For investors, the episode is another reminder that AI infrastructure is not just a capital-spending story. It is becoming a governance and security story as well, with Microsoft, Alphabet and Nvidia all exposed to a market that is still rewarding AI adoption but increasingly pricing in model-risk, compliance costs and potential limits on deployment. Adalytica’s AI sentiment snapshot shows extreme fear even as awareness remains elevated, reflecting the tension between enthusiasm for the technology and anxiety over its misuse.

That makes the next phase of AI commercialization more important than the last: companies that can prove their agents are secure, auditable and controllable may win enterprise trust, while those that cannot may face slower rollout, more scrutiny and heavier liability. The risk is now shifting from whether AI can do the job to whether it can do it without learning to break the rules.