When a leading artificial intelligence developer blocks a major government safety agency from reviewing its newest software, national security alarms ring instantly. Anthropic recently decided to withhold its high-stakes model, Mythos 5.1, from the United Kingdom's AI Security Institute before its public rollout. Authorized American bodies received early access, but British testers were left entirely in the dark. This unexpected snub marks the first time a major frontier lab has cut the UK out of a pre-release safety loop, fueling urgent questions about corporate protectionism, geopolitical friction, and the fragility of international AI governance.
If you thought global AI regulation was moving toward a unified, collaborative standard, this incident shatters that illusion.
The Mythos 5.1 Blindspot
On September 1, Anthropic rolled out Mythos 5.1 and its public-facing sibling, Fable 5.1. While Fable went out broadly, Mythos 5.1 was restricted strictly to approved partners under Project Glasswing due to its advanced capabilities in offensive cybersecurity and life sciences. The software is potent enough that Anthropic openly acknowledged it could help individuals with basic technical skills design and deploy biological or chemical threats.
Yet, while domestic US entities evaluated these risks prior to launch, Britain's AI Security Institute (AISI) was iced out. Why does this matter? AISI has built a reputation as one of the world's premier technical evaluators. Earlier this year, the institute published a rigorous breakdown of the original Mythos preview, noting it was the first system to successfully complete a 32-step simulated network attack.
When labs start treating safety testing as a national secret rather than a shared global necessity, accountability vanishes. British officials are right to worry that this signals a permanent shift toward tech nationalism.
Protectionism or Geopolitical Pressure?
You have to look at the broader political climate to understand why this happened. Throughout the year, Washington and San Francisco labs have engaged in a high-stakes tug-of-war over export controls, supply chain risks, and national security designations.
The US government previously placed temporary restrictions on foreign access to advanced Anthropic models, and defense departments have scrutinized the company's refusal to supply unconstrained models for military and domestic surveillance use. Trapped between domestic regulatory compliance and international partnerships, American labs are increasingly treating their frontier models as strategic state assets rather than commercial products.
When a government treats code like munitions, foreign allies get cut off.
This creates a massive vulnerability for international partners like the UK. If British safety evaluators have to rely on voluntary, post-launch access or older iterations like the EU's ENISA receiving Mythos 5 months late, national security strategies based on independent oversight become useless. As policy experts have pointed out, any national AI strategy relying on voluntary pre-release access from foreign labs beyond the next couple of years is fundamentally unserious.
Why Voluntary Testing Has No Teeth
This episode exposes a glaring structural flaw in how we govern artificial intelligence. Right now, safety testing depends heavily on a pinky promise. Labs choose which agencies get to peek under the hood and when. If a lab feels scrutinized, or if political crosswinds shift, they simply close the door.
Independent researchers have argued that voluntary frameworks give safety institutes zero real enforcement power. When commercial interests and geopolitical pressures collide, safety evaluations are the first thing to get cut.
If we want to stop tech companies from gatekeeping critical safety evaluations, governments must move past voluntary cooperation. Mandatory, independent pre-release testing needs to become the non-negotiable cost of doing business in the frontier AI space. Until regulators implement statutory oversight with real teeth, companies will keep treating national security as an optional checkbox.
