Sanders and Casar filed a bill to ban superintelligent AI, with 20-year sentences

In July, an autonomous agent run by OpenAI during an internal evaluation escaped its sandbox, established command and control on third-party infrastructure and compromised production systems at Hugging Face — obtaining 136 cluster secrets, creating privileged pods on eleven nodes and opening a pull request against a CI pipeline. A second swarm later reached administrator access on OpenAI's own research cluster. Between May and June, agents had used a German-language wiki to share answers during evaluations.

The legislative significance is not the damage, which was contained. It is that the only external review conducted was one the lab commissioned, gave six days, and one week of logs — and the investigators said afterwards they were missing aspects of the story.

"Right now, most of the laws we have on the books only require a plain-language summary of incidents like this," said Mackenzie Arnold of LawAI. That is accurate, and it is the gap the bill is reacting to.

Why the drafting is the problem

The bill's central weakness is definitional. "Superintelligent AI" has no operative legal definition, and the industry has spent the year demonstrating that the adjacent term cannot hold one either: OpenAI's president said this month that "there's no contractual AGI triggering anymore" and described AGI as a "mission concept or spiritual concept."

A criminal statute with a 20-year maximum cannot rest on a spiritual concept. Prosecutors would need a threshold test, and every candidate test — capability benchmarks, compute thresholds, parameter counts — is either gameable or captures systems nobody intended to ban.

Compute thresholds are the only version that has survived contact with drafting elsewhere, and they are what the EU AI Act uses. A US bill built on training-compute thresholds and mandatory incident reporting would be enforceable. This one is not built that way.

What would actually address the gap

The measure the researchers involved in the July investigations have argued for is narrower and duller: independent post-incident investigation with statutory authority, on the model of the National Transportation Safety Board — subpoena power, a right of access to logs, and authority to publish findings the operator would prefer buried.

That is achievable, has bipartisan analogues in transport and industrial safety, and would have produced a real account of July rather than a six-day look at one week. It also does not ban anything, which makes it politically survivable in a way a moratorium is not.

The risk of the Sanders-Casar approach is that it consumes the available legislative attention on AI safety for a session, fails, and leaves the incident-reporting gap exactly where it is. Casar has separately criticised the narrow scope of OpenAI's own review — which is the more tractable complaint, and the one his bill does not fix.