Anthropic's FAA-style AI proposal lets regulators block frontier releases
Anthropic CEO Dario Amodei published an essay calling for FAA-style regulation of frontier AI models, with proposed thresholds that would apply to Anthropic's own frontier releases. According to the source, the proposal would require third-party testing for models trained above 10^25 floating-point operations, or developed by companies exceeding $500 million in AI revenue or $1 billion in AI R&D spending, and would give regulators authority to block or reverse deployment. The framing arrives alongside Anthropic's release of Claude Mythos 5, a gated model the source describes as offering both defensive and offensive cyber capabilities, and Claude Fable 5, presented as the company's most capable general-release model.
The proposal creates a structural tension: it asks for legal authority to constrain the same category of products Anthropic currently ships. The FAA analogy the source attributes to Amodei states that frontier models, "like airplanes, should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety." That is a pre-deployment certification model, not a post-hoc enforcement one. Treating model release as a regulated event rather than a private commercial decision implies regulators would set the release schedule for any frontier system that crosses the threshold.
The threshold itself determines which current systems would fall under the proposed regime. According to the source, the 10^25 FLOPs cutoff, combined with the revenue and R&D triggers, would capture not just hypothetical future systems but the current frontier class. This suggests the regime's first practical tests would likely be systems already on the market rather than systems not yet built. Anthropic's own Claude Mythos 5, which the source describes as the company's gated, updated model with advanced cyber capabilities, is the type of system the proposal targets. The proposal therefore creates an unusual vendor posture: a model developer publicly advocating for legal authority to block its own releases. The source does not address how this authority would be funded, staffed, or kept independent of the companies it regulates.
The cybersecurity framing in the policy proposal is grounded in the source's description of Claude Mythos Preview's ability to discover high-severity vulnerabilities across major operating systems. According to the source, Amodei describes this capability as having "scrambled" the global cybersecurity landscape. The proposed regime would require frontier developers to protect model weights from both external attackers and insider threats, and to report model distillation attacks where competitors or bad actors train cheaper clones on a primary model's outputs. The source does not specify how insider-threat controls at frontier labs would differ from existing classified-computing standards or how reporting requirements would handle dual-use disclosures without exposing the vulnerabilities the model itself can find.
The economic framework is the proposal's third strand, and the source frames it explicitly as a labor-displacement plan rather than a productivity-optimization one. The Economic Policy Framework plans for scenarios where AI drives unemployment to 5%, 10%, or higher levels, and proposes wage insurance, universal basic income, and sovereign wealth fund mechanisms as responses. Anthropic is committing $350 million to back this: $200 million for an Economic Futures Research Fund and $150 million for a national fellowship program. The source attributes a quote to Amodei that frames the central challenge as distributing AI's gains rather than producing them, which is a different policy premise than productivity-focused AI advocacy. The source does not specify which of these mechanisms Anthropic considers most likely or how the fellowship program relates to the research fund in operational terms.
Supply-chain volatility is the proposal's most concrete procurement consequence. The source presents the regulatory regime as one where a flagship model update could be blocked indefinitely, or where an existing model could be revoked if post-release testing reveals autonomous threats. The source recommends multi-model architectures that avoid locking into a single vendor, framing this as a business-continuity requirement rather than a redundancy preference. The source does not specify how procurement contracts would handle revoked-model scenarios, whether liability would shift to the regulator or remain with the developer, or what disclosure obligations would attach to a model under review.
The proposal remains a proposal. The source describes a regime Anthropic is publicly advocating for, not one that has been introduced as legislation or adopted by any regulator. The FAA analogy is the source's framing, not an established policy template, and the 10^25 FLOPs threshold is a proposal Anthropic has put forward rather than a defined regulatory cutoff. Whether the policy architecture the source describes would actually apply to the Claude Mythos 5 release Anthropic shipped alongside the essay, or to future Claude Fable-class general releases, depends on legislation and rulemaking the source does not describe.