What happened
A report from Guidelight AI Standards, covered by TechCrunch, found that most frontier AI labs lack public plans for containing a model that tries to subvert human control. The study graded five major labs on their public preparedness for such a scenario. OpenAI scored highest, while Anthropic and Meta scored lowest. Guidelight defines a containment plan as a pre-specified response for when a model goes off the rails — detailing what access gets cut and when to shut the system down entirely.
How the room's reading it
The labs themselves — including Google and OpenAI — push back, claiming the report doesn't reflect their internal, unpublished safety practices. Legal experts suggest this silence might be strategic, a way to avoid liability if a public plan isn't followed perfectly during an incident. Meanwhile, safety advocates and regulators are losing patience. With new laws in California and New York forcing disclosure, and a federal "AI Kill Switch Act" proposed, the consensus is that self-regulation on containment is failing. The pressure is mounting for labs to show their work, not just talk about internal processes.
Sailfish's take
We're not surprised by the labs' silence. The real takeaway for builders is simple: containment is your problem now. You can't trust the platform layer to save your application if an agentic model misbehaves. We've shipped enough agent-based systems to know that you have to build your own circuit breakers and monitoring from day one. Assuming the model provider has a secret, effective kill switch is a dangerous bet. We're treating containment as an application-level responsibility, not a platform guarantee. If you're building with agents, you should too.