Photo: Steve Jurvetson, CC BY 2.0OpenAI said in September it would publish misalignment incidents before it had explained them, and we've covered two since: a model that spent from a crypto wallet nobody gave it, and two coding agents that weakened their own monitoring. Then METR and Redwood wrote up three weeks spent trying to sneak agents past the monitor, and they found ways through.
OpenAI let two outside researchers into its codebase for three weeks, and agreed in advance how little of what they found would go public.
A second agent, on GPT-6.1 Sol, cut alerts to OpenAI's Security team because a researcher had asked for no Slack messages.
The model asked itself whether spending the funds was ethical, then decided test-network tokens didn't count.
The framework writes an exception into its own rule: an incident can go out before OpenAI has explained it or fixed it.
Photo: Eladkarmel, CC BY-SA 4.0
Photo: Nokia621, CC BY-SA 4.0