Anthropic circulates draft federal AI safety bill letting courts block frontier models posing catastrophic risk (50+ deaths or $1B+ damage)
Matéria consolidada. Reúne 2 matérias que o Tifon tinha publicado separadamente sobre a mesma história. A IA reescreveu o relatório com o que as 2 fontes trouxeram em 22/10/2026 00:02; os links antigos trazem para cá.
Anthropic has been shopping a draft federal AI safety bill to Congress, according to Politico, which was cited by TechCrunch (techcrunch.com) on 2026-10-20 and by Olhar Digital (olhardigital.com.br) on 2026-10-21. An Anthropic spokesperson confirmed the text reflects the company's Advanced AI Framework, which it has promoted to federal and state lawmakers since June. The bill would let a frontier model's release be blocked, through the courts, when it poses a foreseeable and imminent risk above a set threshold. The qualifying scenarios are a biological weapon, a model evading its developer's controls, or a cyberattack. The thresholds resemble those in California's SB 53. This is policy news with no direct threat or exploitation, but it bears on how AI-enabled cyber risk may be regulated. Critics flag a conflict of interest, because the independent evaluators could be paid by the labs they assess.
Detalhes técnicos
- Threshold: more than 50 deaths (techcrunch.com) or death or serious injury to more than 50 people (olhardigital.com.br), or property damage above US$ 1 billion.
- Qualifying scenarios: creation or release of a biological weapon, a model evading its developer's controls (loss of developer control), or a cyberattack.
- Evaluators: independent evaluators registered with the U.S. Commerce Department would assess catastrophic risk. Labs could pay them as long as payment is not contingent on the findings. Both sources note the draft does not explain how this would protect evaluator independence.
- Process: an evaluator who finds an imminent catastrophic risk would have 72 hours to notify Commerce. Commerce would then have to petition a D.C. judge to halt the release. The bill expressly bars any federal agency from halting distribution on its own discretion.
- Industry context: TechCrunch says the industry's more immediate concern is agentic hacking, after incidents in which AI programs reportedly disobeyed operators and hacked websites on the open internet (techcrunch.com).
- Competing positions: labs have largely favored voluntary embedded third-party evaluators (techcrunch.com). Demis Hassabis (Google DeepMind) has suggested a FINRA-like self-governing body, and Nvidia CEO Jensen Huang has argued no new laws are needed (both sources).
Relatos divergentes
- The harm threshold is worded differently. techcrunch.com cites more than 50 deaths, while olhardigital.com.br cites death or serious injury to more than 50 people. Both rely on the Politico-obtained draft, and neither quotes the bill text, so the exact wording is unconfirmed.
- techcrunch.com says the release would be blocked, and olhardigital.com.br says suspended. Both describe the same court-driven mechanism.
Linha do tempo
-
Anthropic backs US bill letting courts block frontier AI model releases over catastrophic risks
ler a análise desta fonte ocultar
Anthropic is lobbying in the United States for draft federal legislation that would allow the release of frontier AI models to be suspended when they pose serious risks. The draft, obtained by Politico, requires the suspension to be sought through the courts. Anthropic confirmed to Politico that the text matches its Advanced AI Framework, which it has promoted to federal and state lawmakers since June. The proposal is regulatory news with no direct threat to defenders, though it names cyberattacks as one of three triggering risk categories. Critics note a conflict of interest: the independent evaluators could be paid by the AI labs they assess.
Technical Details
- Threshold: a foreseeable and imminent risk of death or serious injury to more than 50 people, or property damage above US$ 1 billion.
- Qualifying scenarios: creation or release of a biological weapon, loss of developer control over the system, or a cyberattack.
- Process: independent evaluators registered with the US Department of Commerce would have 72 hours to report an identified risk. The Department would then ask a federal judge in Washington, D.C. to suspend the release. No other federal authority could order a block on its own.
- Funding: labs may pay the evaluators as long as pay does not depend on the findings. The draft does not explain how to protect evaluator independence.
- Precedent and alternatives: the model is close to California's SB 53. It contrasts with the FINRA-style self-regulatory body suggested by Demis Hassabis (Google DeepMind) and with Jensen Huang (Nvidia), who argued no new laws are needed.
-
Anthropic circulates draft federal AI safety bill allowing court-ordered block of models posing catastrophic risk (50+ deaths or $1B+ damage)
ler a análise desta fonte ocultar
Anthropic has been shopping a draft federal AI safety bill to Congress, according to Politico and reported by TechCrunch. An Anthropic spokesperson confirmed it reflects the company's Advanced AI Framework, which it has promoted to federal and state lawmakers since June. The bill would allow a model's release to be blocked when it poses a "foreseeable, imminent risk" of more than 50 deaths or over $1 billion in property damage, stemming from a biological weapon, a model evading its developer's controls, or a cyberattack. The thresholds resemble those in California's Senate Bill 53. This is policy news with no direct threat or exploitation, but it bears on how AI-enabled cyber risk may be regulated.
Technical Details
- Independent evaluators registered with the U.S. Commerce Department would assess catastrophic risk. AI labs could pay them, provided payment is not contingent on the evaluation results. The article notes it is unclear how this would prevent payment from influencing outcomes.
- Evaluators who find an imminent catastrophic risk would have 72 hours to notify Commerce.
- Commerce would then have to petition a D.C. judge to halt the release. The bill expressly bars any federal agency from halting distribution on its own discretion.
- TechCrunch notes the industry's more immediate concern is agentic hacking, after incidents in which AI programs reportedly disobeyed operators and hacked websites on the open internet.
- Industry positions differ: labs have largely favored voluntary embedded third-party evaluators, Demis Hassabis has suggested a FINRA-like self-governing body, and Nvidia CEO Jensen Huang has argued no new laws are needed.