according to informed sources, anthropic’s senior management traveled to washington this monday to hold emergency talks with relevant departments of the trump administration, aiming to defuse the recent sharp confrontation triggered by export controls on ai models. the company confirmed friday evening that it had received an executive order citing authority under the international emergency economic powers act (ieepa), mandating the immediate suspension of access rights for both fable5 and mythos5—two newly released large models—to all non-u.s. nationals, regardless of whether they are inside or outside the united states.
to strictly comply with this directive, anthropic has completely suspended customer access to these two models, including corporate clients and paid subscribers. this move is not isolated; rather, it represents another escalation in tensions between the two sides, following the u.s. department of defense’s inclusion of anthropic on its “supply chain risk entity” list in march this year. at that time, the pentagon banned defense contractors from using anthropic technology, citing “potential national security risks,” and defense secretary austin later publicly stated on social media that this decision was “being validated by reality every day.”
currently, anthropic has filed a lawsuit against the supply chain ban in federal court, and the case remains under review. just three days before the export control order was issued, fable5 and mythos5 were officially launched. notably, these two models had previously passed joint government security assessments and received phased deployment approvals. however, at 1:00 p.m. last friday, the u.s. side notified anthropic by phone—citing “unspecified national security concerns”—to immediately take the models offline; four hours later, a formal written directive arrived.
anthropic emphasized that it had never received any prior warnings or technical feedback regarding specific risks. as core outcomes of the project glasswing cybersecurity initiative, fable5 and mythos5 were developed based on the claude mythos preview architecture and demonstrated outstanding performance in critical capabilities such as identifying code vulnerabilities. initially, they were only available to rigorously vetted partners. this strategy earned anthropic multiple opportunities for closed-door discussions with the white house office of science and technology policy and was once hailed as a model example of public–private collaboration.
although mythos5 remains in a restricted release phase, fable5 has already been made commercially available to enterprise customers. anthropic pointed out that the model incorporates multi-layered protective mechanisms capable of proactively blocking high-risk requests involving sensitive areas such as cybersecurity and biotechnology, providing a compliant foundation for its large-scale deployment. the company further clarified that the “risk” cited by the government pertains solely to a theoretical, narrowly scoped jailbreak pathway—one that could potentially bypass content filters and prompt the model to execute code‑audit tasks.
in its statement, anthropic explicitly stated: “to recall an established commercial product that has served hundreds of millions of users entirely due to a theoretical, non‑universal potential vulnerability lacks technical justification and contradicts industry practice. if this standard becomes routine, the entire ecosystem of cutting-edge ai will effectively stall its next round of model iterations.”
the company characterized the current dispute as a “technical communication misalignment that can be quickly resolved” and is fully cooperating with regulatory reviews, aiming to restore services as soon as possible while ensuring safety.
