White House Blocks Anthropic's Mythos Access Expansion

According to The Wall Street Journal, the White House informed Anthropic that it opposes the company's plan to expand access to its advanced cybersecurity model Mythos to roughly 70 additional companies and organisations, which would have raised the total number of authorised users to about 120. The WSJ reports administration officials cited security concerns and, in some cases, worried that Anthropic would not have enough computing capacity to support a substantially larger user base without affecting government access. The Next Web, citing Bloomberg reporting, adds that the National Security Agency is among government agencies currently using Mythos and that a recent breach during the model's limited rollout intensified administration anxiety.

Per public reporting in The Next Web and Bloomberg, Mythos is an AI system Anthropic developed to find and exploit software vulnerabilities; testing disclosed by outlets indicates the model autonomously discovered thousands of zero-day vulnerabilities across major operating systems and web browsers and achieved 73% success on expert-level capture-the-flag cybersecurity tasks. Those performance claims and the model's autonomous offensive capability are described in coverage of the controlled rollout; reporting also notes Anthropic framed Mythos as too risky for general release and therefore limited initial access.

Industry observers note that tools demonstrating autonomous vulnerability discovery are qualitatively different from conventional code or assistant models because they combine offensive capability with scale, increasing both defensive utility and misuse risk. Companies operating similar systems have faced a tension between enabling broad security research and containing operational risk, and the compute-resource question reported by the WSJ highlights a practical constraint: shared, high-throughput inference for a powerful model can create contention between private testers and government users.

Editorial analysis: Government pushback on a commercial model's rollout places this story at the intersection of AI governance, national security, and vendor operational limits. Reporting by major outlets indicates the episode is not merely a commercial dispute but part of broader, ongoing frictions between Anthropic and federal authorities that have previously covered defence-related concerns. For practitioners, the episode underscores how access policies, infrastructure allocation, and external oversight can directly shape who can use high-risk models and under what constraints.

Observers should track whether the administration pursues formal measures beyond expressed opposition. Benzinga and other reporting say the White House has considered executive workarounds to secure government access; monitor for any draft executive orders, procurement changes, or classified agreements that would alter access. Also watch for clarified third-party audit, vulnerability-disclosure, and compute-provisioning arrangements from Anthropic or prospective participants in any expanded access program. Finally, follow technical publications or red-team reports that either corroborate or contextualise the Mythos testing metrics (thousands of discovered zero-days, 73% capture-the-flag performance) so practitioners can assess real-world risk and defensive value.

The Wall Street Journal is the primary source for the White House objection and the described compute and security concerns. The Next Web and Bloomberg reporting provide additional detail on Mythos capabilities and the involvement of agencies such as the National Security Agency. Different outlets report different company reactions: WSJ and TNW say Anthropic declined to comment, while the New York Post published a quoted company spokesperson; those discrepancies are reported here as differences in coverage.