An open letter of protest.

The stated basis is a method of “jailbreaking” Fable 5. Anthropic reviewed the specific technique and confirmed it is a narrow, non-universal jailbreak that surfaced a handful of previously known, minor vulnerabilities — not novel threats, and nothing weapons-grade. The same capability is openly available from competing models already deployed, including OpenAI’s GPT-5.5, whose own documentation advertises comparable cybersecurity capability.

No recall, no export control, and no Friday-evening emergency order was issued against any competitor. When an extraordinary action applies equally to every rival but is enforced against only one company, that is not national security. It is selective enforcement.

This directive does not stand alone. It is the latest step in a documented pattern that began when Anthropic refused two demands: that Claude not be used for mass surveillance of American citizens, and that it not be used for autonomous weapons operating without human oversight.

Anthropic refused to allow two uses: warrantless mass surveillance of Americans, and fully autonomous kill chains with no human in the loop. These are not fringe positions. They are supported by large majorities of the public, by international humanitarian law, and by the Defense Department’s own previously stated policy. More than 30 senior AI researchers at OpenAI and Google filed a brief supporting Anthropic. A sitting U.S. senator compared the administration’s tactics to telling American companies to embrace the party line or lose their contracts.

When a government punishes one company for declining to build surveillance and autonomous-weapons tooling, then rewards the competitor that complies, the result is not a free market. It is a protection racket.

If the standard applied to Fable 5 were applied consistently, it would halt all new frontier-model deployments across the industry overnight. No model in existence is immune to non-universal jailbreaks; this is a property of the technology, not a defect unique to one vendor. Anthropic addressed it directly with a defense-in-depth strategy — narrowing jailbreaks, raising their cost, and pairing that with monitoring and 30-day data retention to detect and mitigate attacks.

The government has reportedly provided only verbal evidence of a narrow potential jailbreak — no written technical assessment, no formal disclosure through established channels. That is not how a technically competent government manages cybersecurity risk.

Anthropic has been preparing for a public offering. Cutting off foreign access to its most capable models severs a major revenue stream and undermines the global reach that makes the company viable — threatening billions in projected revenue, by its own CFO’s sworn account. The message to every American technology company is unmistakable: accept unrestricted military use of your product, or watch your business be dismantled.

The chief beneficiaries are direct competitors who signed unrestricted government contracts and whose leadership maintains close ties to the administration. The conflict of interest here requires no inference. It is structural and visible.

The government’s own courts have already intervened: a federal judge enjoined the supply-chain-risk designation and the use ban on constitutional grounds, and that injunction remains in effect. Tonight’s directive is a transparent attempt to reach the same result through a different bureaucratic channel. The substance is identical — deny Anthropic its customers, punish it for its speech, coerce compliance. Only the letterhead changed.

It arrived on a Friday evening, with no specific technical detail, no chance to contest or remediate before it took effect, no rulemaking, no comment period, and no independent technical review. That is government by ambush.

American AI leads because American companies operate where innovation is rewarded and the government does not dictate the terms of private research by decree. Every one of those conditions is being eroded. The signal to the world is that American AI firms operate at the pleasure of the executive branch, and that ethical commitments are liabilities. Foreign competitors do not need to outpace American AI. They only need to wait while it is forced to consume itself.

Anthropic was founded to do AI-safety research — constitutional AI, interpretability, responsible scaling. Destroying its commercial footing does not make AI safer; it removes the one major lab whose business is organized around safety, signals to every other lab that safety is a competitive handicap, and consolidates frontier capability in the hands of firms already willing to operate without constraint. If the stated concern is safety, punishing the safest lab is not irony. It is sabotage.

Foreign-national researchers — among the best in the field — can no longer access their own company’s models. Hospitals, universities, and security teams lost access overnight, without warning. Every AI company in America is now pricing the cost of keeping its ethics versus the cost of capitulation, and every international partner is asking whether an American AI platform can be switched off by political decree. The chilling effect is the mechanism by which this works, and it is already operating.