Amodei thanks critic Calvin as Anthropic rewrites AI safety bill

Dario Amodei thanked critic Nathan Calvin on Wednesday for scrutinizing Anthropic, days after Calvin helped drive a backlash against the company’s leaked AI safety bill. Anthropic is just weeks away from the launch of its planned IPO and has responded to the backlash by promising to rewrite parts of the bill. What is the AEGIS Act?

In a post on X, Dario Amodei said that the criticisms by Nathan Calvin, who is the general counsel at the advocacy group Encode, on the leaked AI safety bill, tentatively called the Artificial Intelligence Evaluation, Governance, and Incident Safety (AEGIS) Act, are coming from “a constructive, honest place. ” He added that the public benefits from having people whose job is to examine AI labs and thanked Calvin. Safety advocates keep comparing the bill to the bipartisan FRONTIER Act from Representatives Jay Obernolte and Lori Trahan, which was introduced in July.

The draft is reportedly weaker than the bipartisan FRONTIER Act in two ways. First is that it sets no standards or licensing for the third-party evaluators who check labs’ models, letting them self-certify, and second is that it makes a dangerous model far harder to block. Charlie Bullock of LawAI warned that without licensing, auditors would undercut one another to offer the lightest checks at the lowest price.

Calvin went further, saying the proposal made him “viscerally angry” and likely leaves a court with less power to block frontier model development than exists under current law. California State Senator Scott Wiener, author of the state’s frontier AI law SB 53, wrote on X that the plan “falls far short of protecting public health and safety. ” Under the draft, the Commerce Department could ask a federal court to pause an AI model, but only if there was a clear, near-term risk that would lead to the death or serious injury of more than 50 people, or more than $1 billion in damages.

The harm must also come from either a bioweapon, a model escaping its maker’s control, or a cyberattack. No other agency will be able to suspend a model. The bill would also cancel out some state rules on disclosure, outside testing and incident reporting, but it leaves child-safety and civil-rights laws in place.

Outside testers, called evaluators, would register with Commerce and labs would pay them, as long as the pay did not depend on the results. Evaluators would have to report an imminent risk to Commerce within 72 hours. Anthony Cimino, who leads Anthropic’s federal affairs team, sorted the complaints about the bill into four groups: preemption, model blocking, risk thresholds and updates.

He admitted the preemption wording was “not precise enough” and promised a rewrite. He said states should play a role in setting evaluator standards and in deciding when a lab must shut down a model. Cimino said the dollar and casualty limits were taken from California law, but agreed it is probably time to review them.

He also noted that the draft already lets a court stop a lab’s internal use of a dangerous model, not just a public release. It also requires a report when a model breaks out and reaches outside systems. Will Anthropic’s controversies affect its IPO?

Anthropic responded to the backlash by listing the expensive safety work it has already paid for. Its Frontier Safety Roadmap, which was updated October 19, says the company finished its internal red-teaming goal ahead of a January 1, 2027 target. An automated red-teaming system it built matched or beat the best human entry in all 24 question-classifier pairings drawn from its biological-weapons bounties, winning 23 outright, and did so in 24 to 48 hours rather than the weeks or months human researchers needed.

The same update describes new work on inference verification, isolated inference environments, datacenter physical security and platform security. Separately, Project Glasswing has committed up to $100 million in usage credits and $4 million in donations to open-source security groups, after an unreleased Anthropic model flagged thousands of zero-day vulnerabilities, including a 27-year-old flaw in OpenBSD. Cryptopolitan reported that the company pledged 5% of its shares to a public benefit fund earlier this month.

In February, however, Anthropic’s RSP v3. 0 dropped the old pause commitment, replacing it with Risk Reports and the non-binding roadmap. Anthropic filed its IPO confidentially with the SEC in June and is chasing a valuation above $2 trillion, with the offering expected after the November U.

S. midterm elections. Its prospectus showed a $42 billion loss in 2025, though about $34 billion of that was an accounting charge, and an operating loss that widened to $8.

06 billion. The filing also lists $518 billion in compute commitments over a decade, roughly 80% of which Anthropic cannot cancel. The smartest crypto minds already read our newsletter.

Want in? Join them.