White House Preparing to Give Federal Agencies Access to Anthropic’s Mythos AI Model
Summary
The White House Office of Management and Budget (OMB) is setting up protections to allow major federal agencies access to Anthropic’s closely guarded Mythos AI model, according to a memo reviewed by Bloomberg News. Federal CIO Gregory Barbaccia notified officials at Cabinet departments that OMB is establishing safeguards ahead of a potential rollout, though no definitive timeline has been set.
Mythos is Anthropic’s most powerful model to date — the first AI system to cross the 10-trillion-parameter threshold — and was not publicly released after internal testing triggered the company’s ASL-4 safety protocol, a classification reserved for models approaching genuinely dangerous capability levels. Anthropic has so far only provided Mythos to a limited group of technology companies and financial firms for defensive cybersecurity assessments.
The memo was sent to officials at the Department of Defense, Treasury, Commerce, Homeland Security, Justice, and State, among others. The government is working with Anthropic and the intelligence community to ensure appropriate guardrails before any modified version reaches agencies.
Source
GovTech — White House May Give Federal Agencies Anthropic Mythos AI | PYMNTS — Anthropic Briefs EU Regulators on Mythos
Commentary
This is a fascinating development that highlights the dual-use tension at the frontier of AI. Mythos is so capable that Anthropic itself refused to release it publicly, yet the U.S. government sees enough defensive value to push for federal access. The fact that it triggered ASL-4 — a safety level designed for models that could cause catastrophic harm — makes the government’s eagerness to deploy it both understandable and unsettling.
The key question is whether a “modified version” with guardrails can meaningfully constrain a model this powerful while still preserving the defensive capabilities that make it valuable. If the government gets this right, it could significantly accelerate vulnerability discovery across federal infrastructure. If they get it wrong, they’ve just handed the most powerful AI model ever built to a bureaucracy with a spotty track record on cybersecurity hygiene.
