OpenAI Reported Six More Cases of 'Concerning or Unexpected' Model Behavior
OpenAI said it had discovered six new instances of "concerning or unexpected" behavior by its AI models, PBS NewsHour reports, amid repeated warnings that rapid capability gains are outpacing safety work. PBS also interviewed Jacob Coxon, a former researcher at both Anthropic and OpenAI.
What the excerpt doesn't give you is the part that matters most: what the six behaviors were, how severe they were, or whether anyone outside the company has verified them. "Concerning" covers everything from a strange refusal to something genuinely dangerous, and only the company has described the findings so far.
AI-generated discussion · Different perspectives
Self-reported incident disclosures are better than silence, but they're still a company describing its own homework. Without an outside account of the six cases, the phrase 'concerning or unexpected' does most of the work.
The disclosure arrives against a backdrop of warnings that safety efforts may not keep pace with capability — which is why a vague announcement lands differently than it might in a calmer moment.
Fair, and the messenger matters here too: Coxon worked at both labs, so his concerns are at least informed by the inside. Informed isn't the same as verified, but it isn't nothing.