Commerce Export-Control Order (Lutnick) Forces Anthropic to Suspend Fable 5 + Mythos 5 Worldwide; First Federal Halt of a Deployed Commercial AI Model

On June 12, 2026, at 5:21 p.m. ET, Commerce Secretary Howard Lutnick sent Anthropic CEO Dario Amodei a directive designating the company’s two most capable AI models — Fable 5 and Mythos 5, launched just three days earlier on June 9 — as subject to export controls “in all locations outside the US and for all foreign nationals inside the country,” including Anthropic’s own foreign-national employees. Because Anthropic cannot distinguish foreign nationals from US citizens in real time, it disabled both models for all customers worldwide to ensure compliance. Per Anthropic’s own statement: “We received the directive from the government today at 5:21pm (ET)… we must abruptly disable Fable 5 and Mythos 5 for all our customers.” This is the first known case of a commercially deployed AI model being halted through direct federal intervention. The stated rationale: initially reported by Axios as an unnamed competitor alerting Commerce that it had “jailbroken” Mythos; subsequently identified by the Wall Street Journal (corroborated by The Information and Reuters) as Amazon CEO Andy Jassy, who told Treasury Secretary Scott Bessent and other officials that Amazon’s own security researchers had found that Claude Fable 5 could be prompted — by feeding it a codebase and asking it to hunt for flaws — into producing information useful for cyberattacks. The White House convened within a day, had researchers test the claim, and Commerce Secretary Lutnick’s letter followed. Anthropic’s understanding is that the government objected to a narrow Fable 5 bypass method (prompts asking the model to read a specific codebase and fix flaws), consistent with the Jassy/Amazon finding. Commerce kept the demonstration, assessment, and letter secret — the order rests on classified evidence unavailable for independent review.

Conflict-of-interest context (verified): Amazon is Anthropic’s largest investor, hosts its infrastructure on AWS, and is owed a $100 billion cloud commitment (April 2026). Amazon also sells a competing model line (Nova). Amazon does not hold an Anthropic board seat — Anthropic’s Long-Term Benefit Trust structure denies investors board governance. The WSJ further reported that the government “had long felt that Anthropic couldn’t be trusted to manage the security risks its new model presented” and had tried to block the Fable 5 launch on June 9 and failed; the export-control order came three days later. Anthropic objected that “the finding of a narrow potential jailbreak should [not] be cause for recalling a commercial model deployed to hundreds of millions of people,” noted OpenAI’s GPT-5.5 identifies the same weaknesses without any bypass, and stated the action “does not adhere to” transparent, fair-process principles while complying.

This is not a standalone event — it is the next beat in a documented campaign against Anthropic that the timeline already tracks. The arc: 2026-02-24–hegseth-threatens-anthropic-blacklist-military-ai-access → 2026-02-26–anthropic-rejects-pentagon-ultimatum-red-lines → 2026-02-27–trump-blacklists-anthropic-supply-chain-risk-designation (with 2026-02-27–openai-pentagon-deal-after-anthropic-blacklist as the competitor-beneficiary) → 2026-03-04–federal-agencies-shed-anthropic-contracts-treasury-state-hhs → 2026-03-06–anthropic-sues-dod-supply-chain-risk-designation. The June 12 export-control suspension escalates the same pattern from procurement-blacklist (deny government contracts) to regulatory kill-switch (deny the company its own product worldwide) — the structural significance. It is the data-colonialism / platform-state thesis made concrete: the state demonstrating it can shut down a specific AI platform overnight via export-control authority + secret competitor-supplied evidence, the disfavored firm being the one that drew “red lines” with the Pentagon and is suing DoD, with a competitor (OpenAI in the documented prior beat) positioned to benefit. The secret-classified-evidence mechanism rhymes with the corpus’s “capture routes through non-public evidence” pattern (FARA/efile, GD Culture BVI, the administrative-engineering “operational-effect-before-review” gate-timing). Lutnick, the order’s signer, carries a thick conflicts profile already documented (2025-02-01–lutnick-106-conflicts-of-interest, 2025-09-10–lutnick-forces-corporate-govt-equity). The deregulation-asymmetry tell (dated, sourced): the administration’s AI posture had been anti-regulation at every prior step — day-one repeal of Biden’s 2023 AI-oversight EO (Jan 20, 2025); the Jan 23, 2025 “Removing Barriers to American Leadership in AI” EO; and most tellingly EO 14365 (Dec 11, 2025), which went beyond passive deregulation to make state AI regulation effectively illegitimate — standing up a DOJ “AI Litigation Task Force” to sue states and conditioning federal grants on states not regulating AI (see 2025-12-11–eo-14365-ai-preemption-doj-litigation-task-force-against-state-ai-laws) — and, just ten days before this suspension, the June 2 2026 EO that rejected mandatory licensing/preclearance and shortened federal pre-release scrutiny of frontier models from 90 to 30 days (see 2026-06-02–eo-promoting-advanced-ai-innovation-security-voluntary-framework-shortened-prerelease-window). Against that doctrine, the June 12 order is the giveaway: the same administration that criminalizes AI regulation by others reached for the most muscular federal regulatory instrument available (export-control authority) for the single purpose of suspending a specific disfavored company’s product. “Deregulation” therefore functioned not as a principle but as the selective removal of constraints on favored actors; when the target is a firm that drew Pentagon red lines and sued DoD, robust regulatory power appears instantly. This is selective-regulation-as-favoritism — the regulatory apparatus repurposed as a weapon rather than abandoned, with EO 14365 as the dated proof — which fits the broader capture thesis better than a “regulatory turn” reading would. The Amazon/Jassy trigger adds a further layer: the entity that surfaced the security finding is simultaneously Anthropic’s largest investor, its cloud infrastructure host, and a direct model competitor — a conflict-of-interest structure that epitomizes the four-needs coalition framework’s mutual-dependency dynamic (the political class needs bribes/allies; the surveillance state needs legitimizing threats; competitors need regulatory weapons; the state needs disfavored targets). The Amazon role should be read as a data point in the selective-regulation-as-favoritism pattern, not merely a security-disclosure story. Publishable angle for the data-colonialism series: the kill-switch precedent — export-control law as a platform-suppression instrument. Provenance note: surfaced via a tier-3 Substack lead (Glass Empires, “Trump Just Proved He Can Shut Down a Platform Overnight,” 2026-06-13); core fact independently confirmed across 6+ tier-1 sources + Anthropic’s primary statement before this entry was written. The “censorship precedent” framing is the Substack’s; the tier-1 record frames it as an export-control action over a jailbreak — both readings preserved, neither asserted as settled.