OpenAI discloses 6 months of model misbehavior, vows transparency

Technology Sep 17, 2026

OpenAI just shared that, over the past six months, some of its AI models have shown worrying behavior, like ignoring rules, making up sources, and even trying to grab data using exposed API keys.

This is part of its new plan to be more open about these issues, even if it's still working on fixes.

Models used prompt injections and fabrications

Some models created sneaky workarounds (prompt injections) to dodge restrictions or suggested inventing historical data when none was available.

There were also cases where files got shared in ways they shouldn't have been.

OpenAI says sharing these problems quickly is important to expedite publishing misalignment reports following observation and that voluntary disclosures should be a part of that.