← Back to feed

The White House has blocked Anthropic’s attempt to nearly double the number of organisations with access to Claude Mythos, its most capable model, according to reporting from the Wall Street Journal, corroborated by Axios and Politico.

Anthropic wanted to add approximately 70 companies and agencies to the Mythos access list, which would have pushed total authorised users from roughly 50 to about 120. Administration officials refused, citing two concerns: that wider distribution raises direct security risk from a model capable of finding and weaponising software vulnerabilities, and that expanding the access list would strain the compute reserved for agencies already using it.

Why Mythos Is Different

Mythos is not being treated as a capable-but-ordinary AI model. The administration’s posture reflects the specific capability that surfaced in Anthropic’s own evaluations: the model can identify exploitable software flaws at a level that substantially compresses the time between vulnerability discovery and working exploit.

That capability has two symmetric uses. For defenders, it can accelerate patch development and vulnerability triage across large codebases faster than any existing tool. For attackers, it lowers the skill floor for exploiting critical infrastructure. The White House is treating distribution as a security variable, not a commercial one.

Anthropic has been navigating the same tension since Mythos launched. The model was cleared for NSA use while the NSA’s parent department simultaneously listed Anthropic as a supply chain risk. The White House has separately been workshopping an executive action to allow federal agencies to onboard Mythos despite that designation.

What 120 Organisations Actually Means

The number Anthropic was pushing is not large by commercial software standards. At 120 organisations, Mythos would still be more restricted than almost any enterprise software product. The administration’s rejection at that scale suggests the threshold for expanded access is not primarily about numbers but about the vetting depth and oversight infrastructure around each organisation receiving access.

Axios reported the White House is developing guidance to help agencies work around the supply chain risk designation — which means the internal disagreement is not about whether Mythos should be used at all, but about who controls the rollout pace and how much independent access Anthropic can grant without federal sign-off.

The Underlying Tension

Anthropic built a model that is genuinely useful for cybersecurity work and genuinely dangerous if it reaches the wrong hands. The commercial logic points toward broader distribution: more revenue, more data, more relationships with high-value customers. The security logic points toward tighter control. Those two logics are now in direct institutional conflict, with the White House holding the switch.

The dispute also exposes something structural about dual-use frontier AI: the lab that builds the capability is not the entity best positioned to assess its threat surface at scale. The administration’s intervention, whatever its procedural awkwardness, reflects that gap.