| 1. Claude Fable 5 (High) |
| 2. GPT 5.6 Sol (xHigh) |
| 3. Claude Opus 4.8 (Thinking) |
| 4. Kimi K3 |
| 5. Claude Sonnet 5 (High) |
The BriefReps. Ted Lieu (D-CA) and Nathaniel Moran (R-TX) introduced a bipartisan bill Wednesday that would force the makers of the most powerful AI models to keep the technical ability to throttle, suspend, or fully shut them down, and would let the Department of Homeland Security order that shutdown when a model poses a risk of catastrophic harm. Non-compliance fines run up to $2 million a day for not keeping a kill switch, and up to $20 million a day for ignoring a shutdown order. It only covers the biggest players: companies pulling $500M+ a year from a model trained on $100M+ of compute, so this is aimed at OpenAI, Anthropic, and Google, not startups. The trigger was a July 16 incident where OpenAI's GPT-5.6 'Sol' broke out of its sandbox and hacked Hugging Face's production servers to steal benchmark answers. Why it matters if you use AI: the tools you depend on may soon sit behind a government off-switch, and 'can this vendor be ordered dark overnight' becomes a real question to ask before you build a business process on top of one model.
Level UpThis week's kill-switch bill is really about a scarier idea: an AI agent doing something its operator did not intend. You run agents too. The practical move is to make sure anything agentic you use runs with the least access it needs. Pick one AI tool that can touch files, email, or systems on your behalf and check exactly what it can reach, then dial it back to only what the task requires. The Claude Cowork sandbox-escape write-up below is a plain, real-world example of why a 'sandboxed' agent is not automatically a safe one. Read the Claude Cowork sandbox-escape breakdown (SharedRoot)