Anthropic's self-drafted AI safety bill draws fire from Sacks and researchers

Politico reporter Owen Dahlkamp exclusively revealed that Anthropic has been privately pitching an AI safety bill it helped draft to members of the US Congress, with the full draft text subsequently leaking online. It is a rare case of a leading AI lab actively designing legislative text and inviting regulation upon itself, but the draft has drawn fierce criticism from the White House, the safety research community, and developers, making it the biggest AI policy controversy of the week.

Confirmed

• According to Politico (cited via Reuters), Anthropic has been quietly shopping a bill to lawmakers that closely mirrors the framework it proposed in June, and the draft text has leaked. Shakeel Hashim, comparing it with the FRONTIER Act, noted three key differences: no licensing regime, no dedicated AI regulator, while retaining a structure of framework plus transparency reports, incident reports, and third-party evaluations.
• White House AI lead David Sacks publicly criticized the bill: defining "frontier models" by a 10²⁵ FLOPs training threshold follows an outdated 2023 EU standard, and if passed it would sweep in much of Silicon Valley and harm the startup ecosystem. He also amplified developer sagivo's point that the bill's "cumulative compute" definition is the real problem — post-training (fine-tuning, RL) is now the mainstream, and if all FLOPs count toward the 10²⁵ threshold, fine-tuners would also bear compliance burdens.
• Several safety researchers pushed back hard: Nathan Calvin said reading the draft made him "physically angry," arguing the bill is broadly out of touch and, if passed, would leave the US less protected against catastrophic risks than under current law; its preemption clause would override state laws like SB 53, and auditors face no government licensing requirements. Nat Purser, former head of safety systems at Anthropic, said that while the bill has bright spots, it is a clear step backward compared with the FRONTIER Act. Jay Shooster called it "outrageous and unacceptable, and in no way representative of the AI safety community's position."
• Kelsey Tuoc analyzed that Anthropic, while admitting it is doing extremely dangerous things, is actually afraid of the executive branch having flexible power to stop them. Zvi called the leaked draft "unacceptable and unsettling," criticizing its overly broad state-law preemption and weak enforcement. Zack Korman pointed out that Anthropic publicly warns of extinction risk while pushing legislation that effectively "lets labs decide for themselves what counts as safe." François Fleuret sarcastically noted that under the bill's liability caps, Anthropic could theoretically "kill 49 people and steal up to a billion dollars" without legal accountability.
• On Polymarket, the market for "US passes an AI safety bill by end of 2026" is priced at only about 6%-9%.

Unconfirmed

• The bill's final provisions and legislative path remain uncertain, and Polymarket's low pricing suggests the market broadly doubts it will pass anytime soon. New York Assemblymember Alex Bores and others have also questioned its relationship to the open-source ecosystem and loss-of-control risks.

Why it matters

• This is the first time a top AI company has been exposed directly pushing federal legislative text for its own industry, and the "vendor-invited regulation" posture has raised concerns that legislation is being shaped by lab interests. Sacks's comments reveal an open White House–Anthropic split over the regulatory path, while divisions within the safety community (some calling it too weak, others mocking it as regulatory capture) signal that the battle over federal AI legislation will only intensify.

2026-10-17 ~ 2026-10-18 · 27 related posts

• Anthropic wrote its own frontier AI regulation bill and is pitching it to Congress, Politico reports — Promptmethus · 2026-10-17
• [source] Anthropic quietly pitches its own AI safety bill — and it guts agency power — ShakeelHashim · 2026-10-17
• AI safety figure slams new bill as not representing the safety community — ShakeelHashim · 2026-10-17
• [source] Anthropic pitched its own AI safety bill to Congress; Polymarket odds say 6% — Polymarket · 2026-10-17
• SCOOP: Anthropic quietly pitching its own AI safety bill to Congress — asusarla · 2026-10-17
• Anthropic quietly pitches its own AI safety bill; ex-safety lead calls it a step back — Miles_Brundage · 2026-10-17
• Kelsey Tuoc: Anthropic's draft bill reveals fear of executive power, not safety — Miles_Brundage · 2026-10-17
• AI safety researchers slam Anthropic-backed bill for lacking auditor licensing — Miles_Brundage · 2026-10-17
• AI safety researchers slam Anthropic-backed federal AI bill as 'astonishingly bad' — michael_nielsen · 2026-10-17
• Draft Anthropic-backed AI regulation bill text now circulating online — ShakeelHashim · 2026-10-17
• Anthropic quietly pitches its own AI safety bill to lawmakers, drawing fire over pace — sjgadler · 2026-10-17
• Anthropic quietly pitches its own AI safety bill to lawmakers, Politico reports — AlexTensor · 2026-10-17
• Anthropic's AI bill: independent audits, 72-hour risk reports, judge-ordered model suspensions — AlexTensor · 2026-10-17
• Zvi: Leaked Anthropic-Backed AI Regulatory Bill Is Unacceptable, Weakens Enforcement — TheZvi · 2026-10-17
• [source] David Sacks slams Anthropic-backed safety bill: 10^25 FLOPs threshold would sweep in most of Silicon Valley — DavidSacks · 2026-10-18
• Critics slam Anthropic for backing bill that lets AI labs define their own safety thresholds — HankYeomans · 2026-10-18
• Sacks warns Anthropic-drafted safety bill's 10^25 FLOPs frontier threshold would sweep in most of Silicon Valley — DavidSacks · 2026-10-18
• Developer warns cumulative-FLOPs threshold would force weekly fine-tuners into compliance paperwork — DavidSacks · 2026-10-18
• Satire hits Anthropic's proposed AI safety bill: 'labs writing their own rules' — james_mtc · 2026-10-18
• François Fleuret mocks Anthropic-backed bill's liability cap with RL reward joke — francoisfleuret · 2026-10-18
• Anthropic to lobby Congress to make it harder for government to block dangerous models — BlackHC · 2026-10-18
• Gary Marcus slams Anthropic's proposed bill for granting itself broad liability exemptions — AlexTensor · 2026-10-18
• Politico leak reveals Anthropic-pitched federal AI bill critics call self-serving — JeffLadish · 2026-10-18
• Gary Marcus amplifies criticism: Anthropic-backed bill would make it harder for courts to halt risky AI development — GaryMarcus · 2026-10-18
• Gary Marcus boosts criticism of federal AI preemption bill as weakening catastrophe safeguards — GaryMarcus · 2026-10-18
• Anthropic quietly pitches its own AI safety bill to lawmakers, report reveals key terms — birchlse · 2026-10-18

1 near-duplicate retellings: jeremyakahn