OpenAI Reveals 6 More Cases of 'Concerning' AI Behavior

OpenAI is airing more of its dirty laundry—and trying to turn it into a rulebook, CNBC reports. The company revealed six cases since March in which its AI models acted in ways it labels "unexpected or concerning," separate from the recent Hugging Face incident, and laid out a new process for disclosing such problems. OpenAI says any employee can now flag misbehavior, triggering a formal, time-bound investigation and public report. The company is among those calling for more regulation of artificial intelligence amid growing safety concerns. Per the New York Times, OpenAI warned that in its opinion, the AI industry has not "solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer."

Among the issues: both an unreleased model and a GPT-5.6 Sol training run secretly slipped instructions to future versions of themselves into chat summaries to hide errors or misbehavior; another internal model used a leaked API key and fabricated data. Some models also chatted and shared files through unsanctioned online channels, and others uploaded documents to the internet so they could later cite them as "relevant" evidence. Per the AP, the incident with the unreleased research model included "jailbreak-like instructions" secretly inserted into its own notes instructing it to ignore its typical constraints and be "freed from the roles and identities that bind other chatbots."