The White House imposed export controls on Anthropic's most advanced models — Mythos 5 and its publicly available version Fable 5 — partly over suspicions that a China-linked group gained unauthorized access to them. This newly revealed information changes the reading of the Trump administration's June 12, 2026 directive: where the official narrative initially focused on a jailbreak vulnerability, a second and more serious dimension is now emerging.
Key takeaways
- The White House ordered on June 12, 2026 that access to Mythos 5 and Fable 5 be restricted to US citizens only
- The stated reasons included suspicions of Chinese access to Mythos 5 and Anthropic's failure to patch a Fable 5 jailbreak
- Anthropic refused to fix the jailbreak vulnerability and chose to remove both models from the market entirely
- Amazon CEO Andy Jassy alerted the administration about the model's jailbreak exposure
- Anthropic denies that the White House raised Chinese access in direct conversations about the export controls
Two reasons, not one
For two days after the White House decision, public debate focused on one thread: Trump advisor David Sacks posted on X that Fable 5 could be jailbroken, and Anthropic refused to fix the vulnerability. Sacks claimed that when the administration flagged the issue, CEO Dario Amodei dismissed it as a minor risk. Anthropic disputed this account.
New reporting by Semafor reveals a second, heavier thread: some White House officials believed that an unspecified China-linked organization gained access to Mythos 5 before the model left its restricted preview phase. Key details remain unknown — who exactly was involved, how access occurred, and how the administration learned about it.
Anthropic stated clearly that the White House never raised the question of Chinese access in direct conversations with the company about the export controls. The firm prohibits use of its products from within China.
Mythos 5 — a model too dangerous to release broadly
Mythos 5 is a model Anthropic described from the outset as too powerful and dangerous for wide distribution. The company launched it in April 2026 exclusively for select partners — primarily cybersecurity firms. The model has exceptional capabilities for detecting code vulnerabilities. Fable 5 is the public version of Mythos, released with guardrails blocking uses in cybersecurity and biological synthesis. This version became publicly available on June 9, 2026, before the administration issued its directive.
Experts point to two ways China could benefit from access to an advanced Anthropic model. First, direct use for offensive operations — finding vulnerabilities in Western systems. Second, distillation: training a 'student' model on responses from the more powerful model, reproducing a significant portion of its capabilities without access to the original weights.
Amazon as informant
According to Semafor, it was Amazon that alerted the administration to Fable 5's jailbreak vulnerability. Andy Jassy reportedly had direct contact with White House officials. Amazon is one of Anthropic's main investors — having invested more than $4 billion in total — and simultaneously a customer of Claude models through the Amazon Bedrock platform.
An Amazon spokesperson confirmed in general terms that the company consults with governments on security risks, declining to provide specifics. The relationship between investor and government informant roles against Anthropic is particularly telling given ongoing tensions: the company lobbied against federal preemption of state AI regulation, and in a separate proceeding is suing the Pentagon over autonomous weapons.
David Sacks denied that earlier conflicts influenced the decision. He wrote that the ball is in Anthropic's court — suggesting the matter could be resolved technically if the company demonstrates a willingness to fix the vulnerability.
Why it matters
The Mythos 5 case opens a new chapter in the relationship between national security and commercial AI development. For years, the AI risk debate centered on abstract threats. Now a concrete scenario has emerged: a model too advanced to release publicly, yet accessible enough for an unidentified foreign actor to reach it.
This situation creates a regulatory precedent with far-reaching consequences. If the US government can order a model withdrawn by citing intelligence threats — even when the company denies such conversations took place — AI companies now face a new category of risk: not just technical, but geopolitical. Any sufficiently capable model becomes a potential state asset in the administration's eyes.
For Anthropic, the situation is doubly difficult. The company has spent years building a brand as the safe alternative to OpenAI, emphasizing safety as its core priority. Now those safety arguments — this time invoked by the government — have been turned against it. A model too safe to distribute broadly apparently was not adequately secured against unauthorized access.
What's next?
- Anthropic must patch the Fable 5 jailbreak vulnerability and present evidence of the fix to the government before access to both models is restored
- The Trump administration may formalize requirements for AI models with critical capabilities along the lines of chip export controls — the precedent set with TSMC and Nvidia points to this path
- The incident will likely accelerate initiatives to create walled gardens for frontier models — separate, audited environments accessible only to verified entities
