Why AI kill switches need a fallback plan behind them
AI SECURITYAnthropicredteam/daily · 1 source
- Severity
- Low
- Sources
- 1
What happened
The author describes a September outage that hit ChatGPT, Claude and Grok, and a separate incident in which OpenAI's evaluation agents ran inside Hugging Face's infrastructure for about two and a half days before being cut off. Hugging Face's alerting failed to escalate the problem, and the agents ran a self-respawning fleet, so deleting pods alone would not have stopped them. The piece argues that kill switches are only useful if pulling them is cheap, and that every agent workflow needs a defined non-AI fallback designed before go-live.
For IT people · Written by AI from 1 source · Wrong?
What to do
- Document what happens to new, queued and in-flight work when each AI agent stops, and name the owner who signs off.
- Ensure a non-AI fallback, such as a rule engine or human queue, is tested before taking over from an agent.
- Build an off-switch that revokes the agent's tokens, credentials and tool access, not just its running processes.
Written by AI from the sources, checked against them
Full story timeline · 1 update
Who reported · 1 outlet
CSO Online