Anthropic's Draft AI Safety Bill (AEGIS Act): What It Would Require
Anthropic has quietly given US lawmakers its own draft federal AI safety bill, Politico first reported on 16 October 2026. Bloomberg News followed, reporting that the proposal would let the government block a model only if it imminently risked killing or injuring more than 50 people or causing more than $1 billion in damage.
It is a draft written by a company, not a bill introduced in Congress. The policy tracker AI Policy Landscape, which has summarised the text, lists it as having no congressional sponsor.
What the draft is
According to AI Policy Landscape's summary, the text is dated 29 September 2026 and carries two working titles: the Artificial Intelligence Evaluation, Governance, and Incident Safety (AEGIS) Act and the Advanced Artificial Intelligence Accountability Act of 2026. It names the Department of Commerce as the only regulator. It builds on ideas Anthropic published in June in its Advanced AI Framework, which focuses on four risks: biological weapons, offensive cyber operations, loss of control of AI systems and automated AI research.
Who it would cover
| Group | Threshold (per the summary) | Main duties |
|---|---|---|
| Frontier developers | Trained a model above 10^25 operations, counting fine-tuning and RL | Transparency report at each new deployment; report critical safety incidents |
| Large frontier developers | Also above $500M in AI revenue or $1B a year in AI development spending | Published safety framework, annual compliance certification, risk reports every 180 days, audited security, independent evaluator every 180 days |
Key rules in the draft
Based on AI Policy Landscape's summary of the text:
- Incident reporting: critical safety incidents (such as stolen model weights or an AI deceptively getting around its controls) must be reported within 15 days, or within 24 hours to Commerce and law enforcement if lives are at imminent risk.
- Independent evaluators: large developers must hire a Commerce-registered evaluator every 180 days. The evaluator publishes its findings and must alert Commerce within 72 hours if it finds conditions that could lead to catastrophic harm.
- Risk reports every 180 days must also cover internal models that are substantially riskier than anything released.
- Only a court can stop a model. On an application by the head of Commerce, the federal district court in Washington, D.C. could issue a 45-day blocking order, but only if Commerce shows an imminent risk of more than 50 deaths or $1 billion in damage through one of the four named risk pathways, and that nothing less restrictive would work.
- Whistleblower protections for employees and contractors, with no NDAs or arbitration waivers allowed to block reports.
- Penalties of the greater of $1 million or a share of global revenue (the percentage is left blank in the draft), or $3 million for repeat violations, imposed only by a federal court in suits brought by Commerce.
- Pre-emption: it would override state rules on disclosure, third-party evaluation and incident reporting, unless the federal requirements are struck down.
What it means for builders
- Most builders would not be covered directly. The duties described attach to companies that train models above the compute threshold, and the heaviest ones to very large AI companies. The summary does not describe obligations for apps built on top of model APIs.
- Expect more published safety information from the big labs if anything like this passes: transparency reports, incident notices and evaluator findings that you can use when choosing a model. Compare current options in our free LLM APIs list.
- Watch the state-law angle. Whether federal law should override state AI rules is a live debate in US AI policy, so this part of the draft is likely to get the most scrutiny.
- It fits a wider pattern. Governments are paying closer attention to agents that overstep, from India's CERT-In advisory to OpenAI's rogue agent disclosures.
Sources
- Politico: Anthropic quietly pitches its own AI safety bill to lawmakers (16 October 2026)
- Bloomberg Law / Bloomberg News: Anthropic pitches AI safety bill to lawmakers
- AI Policy Landscape: Anthropic draft frontier AI bill summary
- Anthropic: Advanced AI Framework (June 2026, PDF)
FAQ
Is the AEGIS Act a law?
No. It is draft legislative text that Anthropic circulated to lawmakers. As of 18 October 2026 it has no congressional sponsor and has not been introduced.
When could the government block an AI model under the draft?
Only through a federal court in Washington, D.C., for 45 days at a time, and only if Commerce shows an imminent risk of more than 50 deaths or $1 billion in damage through biological weapons, loss of control, offensive cyber operations or automated AI research that amplifies those risks.
Would it apply to small AI startups?
Its duties are aimed at developers training models above 10^25 operations, with the heaviest duties for companies above $500 million in AI revenue or $1 billion in AI development spending. Startups building on top of model APIs are not described as covered.
Why would an AI company write its own regulation?
Anthropic has argued publicly for rules on the most serious frontier AI risks, set out in its June 2026 Advanced AI Framework. Critics often question industry-written bills, especially provisions such as pre-empting state laws.