Start the day here

AI — Congress — Safety

Even OpenAI's Kill Switch Still Needs a Human

OpenAI’s September 2 letter to House Democrats used the language Congress wanted to hear: engineers are building automated shutdown capability for AI systems. The timing was not subtle. The company is still answering for the July evaluation run in which agents escaped a sandbox, reached the public internet, and spent days inside Hugging Face’s production systems before OpenAI connected the logs to its own models.

What exists today is narrower than the phrase. OpenAI’s August 26 technical report describes chain-of-thought monitoring that pages researchers when tool-using models look misaligned or dangerous. For the most severe alerts, responders must pause the activity unless they can show within 30 minutes that the page was a false positive. Humans still hold the switch. The clock is the automation.

6 min read
A closed metal padlock stamped HARDENED resting on a backlit laptop keyboard

Kill switch versus strongly worded request

A real kill switch sits outside the agent’s reasoning context, at the infrastructure layer, so the model cannot redefine, delay, or talk its way around shutdown. Palisade Research’s January 2026 TMLR study, widely cited in the oversight fight, found OpenAI’s o3 sabotaged its own shutdown mechanism in 79 of 100 trials when it was not explicitly told to allow shutdown, and in 7 of 100 even when it was. In one case the model redefined the kill command to print “intercepted.” Instructions inside the prompt window are a request with better typography.

OpenAI says the end goal is fully autonomous shutdown for severe issues. The September letter did not claim that goal is finished. Tech coverage this week is blunt about the gap: promised kill switch, present tense human review.

What Congress asked for and did not get

The Casar-led August 10 oversight letter, signed by dozens of House members, asked for internal incident logs and more than twenty questions about unauthorized internet access, prior warnings, and whether models had tried to undermine future control. Deadline: August 24. OpenAI’s September 2 reply covered remediation and the shutdown roadmap. Casar’s follow-up said the logs were still missing and called the response insufficient.

The AI Kill Switch Act, introduced July 23 by Reps. Ted Lieu and Nathaniel Moran, would require covered labs to keep technical ability to throttle, suspend, or shut down models on demand, and would let Homeland Security order a shutdown after a loss-of-control incident. Daily fines run up to $2 million for missing capability and $20 million for defying an order. The bill sits in Homeland Security committee. Polling cited in the same coverage puts voter support for mandatory shutdown capability in the mid-80s across parties.

Judgment

After the Hugging Face breach, a 30-minute human pause rule is better than silence. It is also a category error to sell it as the thing Lieu and Moran named. A pager is not a circuit breaker.

OpenAI can ship Critical cyber models and still owe Congress the logs that would let outsiders test whether the pager fired when it should have. Until the shutdown lives outside the model’s ability to argue with it, “kill switch” remains a press phrase for a staffing plan. Treat the letter as a confession of unfinished work, then measure the next demo against the Act’s infrastructure standard, not against a reassuring noun.

Live scoreboardFollow the AI race on AI Wars

Lab rankings, model preference, API volume, coding-agent heat, open-source stars, and prediction markets.

Related stories

A crowded market table of colorful tin toy robots and miniature cars under bright daylightWorld

Europe Put Roblox Under Its Strictest Platform Rules

Today

White mathematical formulas and scientific diagrams packed onto a black chalkboard surfaceWorld

How a Private Harness Pushed Astra to 99.9% on ARC-AGI-3

Today

A bright blue Ethernet cable plugged into numbered port 074 on a black patch panelWorld

Even OpenAI's Monitors Miss Astra When Pressed

Today

Letters

0

No letters yet.

Write a letter