The Cybersecurity and Infrastructure Security Agency now has full access to Anthropic’s Mythos Preview model, but the agency lacks clear White House guidance on how to use it — a gap that puts a powerful defensive tool into operational hands without settled rules for its employment.
CISA gains full access to Anthropic’s Mythos Preview
Two people familiar with the matter — including a U.S. official who spoke on the condition of anonymity — told Defense One that CISA received full access to Anthropic’s Mythos Preview around a week ago. Both sources discussed internal deliberations and said the agency’s access is new. CISA did not respond to a request for comment.
ONCD has not set clear parameters for model use
The White House Office of the National Cyber Director has not yet set clear parameters for how CISA should use Mythos Preview, the U.S. official said. That absence of formal guidance echoes prior reporting in Nextgov/FCW that federal tech leaders have privately complained ONCD had not adequately briefed them on implementing or using the model for vulnerability scanning.
Project Glasswing rollout, export controls, and recent timeline
Anthropic has been rolling out Mythos Preview through a non-public, selective program called Project Glasswing. Over the last few months the company “surgically” distributed the model to select organizations and expanded the effort to partners in industry and other nations, according to the sources. The model and its successor versions have been restricted on the grounds that, in the wrong hands, they can significantly boost adversaries’ hacking capabilities.
Media accounts provide a shorthand timeline: Axios reported in April that CISA had not been included in an initial Mythos rollout; Nextgov/FCW reported last week that agency access to the model was imminent; and Defense One’s sources say that access arrived about a week ago. Separately, the U.S. effectively banned Anthropic’s Mythos 5 successor model via an export control mechanism over the weekend, alongside the company’s Fable 5 model — a move that has caused uproar across the cyber and AI community. Both Mythos 5 and Mythos Preview have only been made available to vetted providers via Project Glasswing.
How Mythos-class models change defensive and offensive cyber operations
The reporting describes a class of models — exemplified by Mythos and its variants — that can rapidly analyze large bodies of software and system data to surface weaknesses and possible attack paths for human defenders to review. For federal agencies, that capability can speed identification of vulnerabilities. Conversely, the same capability can accelerate offensive hacking operations: the sources say cyber operators in the intelligence community and the Defense Department can use such models to speed attackers’ work as well.
Anthropic’s guarded rollout reflects that dual-use character. Project Glasswing’s vetting is intended to limit distribution precisely because, if misused, the technology could “significantly boost adversaries’ hacking capabilities,” the sources said.
What this means for federal tech leaders, security teams, and adversaries
- Federal tech leaders: Several federal technology officials have privately complained that ONCD has not sufficiently briefed them on how to implement or use Mythos-style models for vulnerability scanning, signaling friction between policy coordination and operational rollout.
- Security teams and technologists: With CISA now holding access, defenders may gain faster tooling for finding and triaging vulnerabilities — provided ONCD and CISA agree on safe operating procedures and limits for deployment.
- Adversaries and threat actors: Anthropic and its government partners are restricting access because the same model that helps defenders can, in the wrong hands, be leveraged to scale offensive operations and accelerate exploitation.
The facts on hand present a pointed operational puzzle: CISA has been given a powerful defensive tool, but the federal playbook for how and when to use it has not been publicly articulated. The story now hinges on whether the White House’s Office of the National Cyber Director will set the parameters CISA needs — and whether Project Glasswing’s vetting and the recent export-control action on Mythos 5 and Fable 5 will be enough to contain dual-use risks as these models proliferate.
