Why did Anthropic deny the UK AI Security Institute (AISI)—which has a £66 million annual budget, more than 100 technical specialists, has evaluated over 30 frontier models, previously tested Claude Mythos 5, and lacks legal power to compel companies to provide models—pre release

Anthropic has not publicly stated why it withheld Mythos 5.1 from AISI, so a definitive motive is unknown. The evidenced fact is narrower: the institute did not receive pre-release access while a small, vetted US group did; the protectionism explanation is an informed concern by UK officials, not a confirmed Anthropic rationale. 34

• What differs between the models: Anthropic describes Mythos 5.1 as its restricted cyber- and biology-research model, available only to selected vetted organisations, while Fable 5.1 is the generally available model for demanding reasoning, coding, research, and long-horizon agentic work. Both have a 1-million-token context window and up to 128,000-token outputs. 124 Anthropic’s earlier description of the same product split was that the restricted model used the same underlying model as the public counterpart but had safeguards lifted in specified high-risk areas. 5

• Why the denial matters: AISI’s role is to independently measure frontier-model capabilities and evaluate mitigations for government. It has over 100 technical staff, but participation is voluntary rather than compelled; therefore, exclusion creates a real blind spot precisely where safeguards have been relaxed. 267

• What cannot be concluded: There is insufficient evidence that the decision was ordered by, or legally required by, the US government. It could reflect export-control caution, national-security sensitivity around cyber/bio capabilities, commercial confidentiality, or a broader US-only access policy. Anthropic’s lack of a public explanation prevents separating these possibilities.

• UK response and geopolitical reading: The government has emphasized that AISI continues to work with frontier developers and has conducted pre-release evaluations of other systems. Nonetheless, officials’ concern is understandable: access arrangements increasingly look like strategic controls rather than a purely global safety-testing norm, especially amid US–China competition and restrictions on highly capable cyber/bio systems. That concern remains an interpretation, not proof of a coordinated protectionist policy. 35

• Safety significance: The concern is not merely procedural. AISI exists to test capabilities that developers may under-measure or that only appear under adversarial evaluation. Reports of increased cyber capability, agentic misuse, a researcher’s resignation, and Evan Hubinger’s personal estimate of more than a 10% existential-risk chance within a decade heighten the case for independent scrutiny—but those statements are warnings and personal assessments, not evidence that Mythos 5.1 itself poses an extinction-level threat. 6

In short, the lack of access does not establish that Mythos 5.1 is unsafe; it means an important independent evaluator could not test whether the restricted, relaxed-safeguard version behaves safely before deployment.