Federal CIOs fault ONCD silence on Mythos access
A cyber model kept dangerous enough to ration needs a distribution plan better than private briefings and agency guesswork.
TL;DR
Nextgov/FCW, citing five people familiar, reports that federal technology leaders are frustrated with the White House Office of the National Cyber Director over unclear plans for accessing and using Anthropic’s Mythos model to scan agency networks. The intelligence community has access, while other CIOs say they lack either access or instructions. That is the adoption gap in one sentence: Washington wants frontier AI for defense, then leaves the operators to learn through private briefings.
Nextgov/FCW’s report is built on anonymous sourcing, so the right read is narrower than a claim that the government cannot use Mythos. Select parts of the U.S. government, including the intelligence community, already have access, and Anthropic has been keeping the model in a non-public distribution channel through Project Glasswing because the same capability that helps find vulnerabilities can help adversaries exploit them. The complaint is operational: CIOs responsible for agency networks say the Office of the National Cyber Director has not told them how access will work, what they may do with it, or whether they may engage Anthropic directly.
That matters because the White House has spent the last month telling agencies to move faster on frontier AI. Nextgov/FCW reported that the June 2 AI executive order encouraged developers to give the government 30 days of pre-release model access and called for a binding operational directive to secure civilian networks and facilitate access to frontier AI for critical infrastructure within 30 days (https://www.nextgov.com/artificial-intelligence/2026/06/trump-signs-ai-executive-order-after-postponement-last-month/413912/). It also reported that a June 8 national security memo told agencies including the FBI, the Office of the Director of National Intelligence and ONCD to build “deep, proactive” relationships with AI companies so models reach national security personnel faster (https://www.nextgov.com/artificial-intelligence/2026/06/trump-memo-pushes-national-security-agencies-move-faster-ai/414031/). Those signals do not answer the mundane question a CIO has to settle before running a model against an agency network: who authorizes it, under what conditions, and with what guardrails.
Anthropic filling the vacuum with federal CIO briefings is useful, and it is also the symptom. Vendors can explain capabilities. They cannot set executive-branch risk tolerance for a model being treated as too dangerous for open release and too useful to ignore. For practitioners, nothing in the report creates a Monday-morning procurement path. It leaves a planning problem: track ONCD guidance, assume access will be uneven, and do not build a vulnerability-management schedule around Mythos until someone with authority says who may use it.
Published ·Deep Fathom