Anthropic withholds Mythos Preview from the public over cyber risks
Restricted access through Project Glasswing prompts debate over cyber danger, marketing and access to defensive tools.
AnthropicAbout these dates
April 7 is the announcement and restricted-preview date. The account includes subsequent criticism; this is separate from the June 9 Fable 5 / Mythos 5 launch.
Restricted release
On April 7, 2026, Anthropic announced Claude Mythos Preview but said it would not make the model generally available. It cited the danger of its ability to find and exploit software vulnerabilities, and said stronger safeguards were needed. Anthropic ↗
Selected organizations instead received access through Project Glasswing for defensive security work. Anthropic's stated aim was to help secure critical software before comparable capabilities reached attackers. Anthropic ↗
Criticism and test limits
AISLE founder Stanislav Fort disputed the idea that effective AI defense depended on a restricted frontier model. His company reported that cheaper open-weight models recovered parts of Mythos's published vulnerability analysis, and argued that the framing risked concentrating defensive capability behind one provider. Stanislav Fort / AISLE ↗
AISLE's tests supplied selected vulnerable code and sometimes hints; they did not reproduce autonomous discovery or exploitation with tools. Its criticism therefore did not establish that those models matched Mythos's full capabilities. Stanislav Fort / AISLE ↗
On April 13, security commentator Bruce Schneier criticized Anthropic's publicity and uncritical coverage while also arguing that AI cyberattack capabilities were advancing. He distinguished finding a vulnerability from turning it into a working attack: questioning the marketing did not mean dismissing the danger. Bruce Schneier ↗
Sources & attribution
- Organizational disclosure 7 Apr 2026Project Glasswing: Securing critical software for the AI era ↗
Anthropic. Original announcement; introduction and general-availability decision below the capability evaluations. Capability claims are Anthropic's.
- Analysis 7 Apr 2026AI Cybersecurity After Mythos: The Jagged Frontier ↗
Stanislav Fort / AISLE. Displayed publication date; the living article includes an April 9 update. See The bigger picture and Caveats and limitations. Interested-party research, not an end-to-end comparison.
- Commentary 13 Apr 2026On Anthropic's Mythos Preview and Project Glasswing ↗
Bruce Schneier. Author's original observations, points One through Four; criticism of publicity alongside acknowledgment of advancing capabilities.