Case mythos-unauthorized-accessModel & data leakUnited States
Unauthorized group accesses Anthropic's restricted Claude Mythos model via vendor
Bloomberg reported a small group reached the restricted cyber-capable Mythos model using a contractor's account and a guessed URL; Anthropic investigated a third-party vendor environment.
HighAttack on AIStatus: Under investigationVerified case file
Open in the interactive mapCite this case
- Reported
- Apr 21, 2026
- Location
- United States
- Target
- Anthropic Claude Mythos Preview
- Severity
- High
- Triggered by
- Small Discord-based group of AI enthusiasts, one with contractor access
- Actor type
- Insider
- Reported loss
- None reported
- CVE
- None
What happened
Bloomberg reported a small group reached the restricted cyber-capable Mythos model using a contractor's account and a guessed URL; Anthropic investigated a third-party vendor environment.
Impact: Unauthorized use of a restricted cyber-capable model by a handful of users over roughly two weeks; no evidence of core-system compromise reported
Who triggered it
Actor: Small Discord-based group of AI enthusiasts, one with contractor access
Attribution: Reported by Bloomberg based on screenshots and live demo; Anthropic confirmed investigating a third-party vendor environment
Motive: Curiosity and experimentation with unreleased models, per the group; no reported malicious use
How it happened
Attack vector: Shared accounts and API keys of an authorized contractor plus a guessed model endpoint URL based on Anthropic naming conventions
Root cause: Weak access controls in a third-party vendor environment: shared credentials and predictable model endpoints with no additional gating
Timeline
- Mythos Preview announced under Project Glasswing; group reportedly gains access the same day
- Bloomberg reports the unauthorized access; Anthropic confirms investigation
- Anthropic says no evidence of impact to core systems beyond the vendor environment
Response
Anthropic investigated the vendor environment; outcome of the investigation not publicly reported
Current status: Under investigation
Sources (open in a new tab)
- The Decoder the-decoder.com
- Security Boulevard securityboulevard.com
- The Next Web thenextweb.com
Cite this case
Unauthorized group accesses Anthropic's restricted Claude Mythos model via vendor. AI Threat Watch, case mythos-unauthorized-access, Apr 21, 2026. https://aithreat.io/cases/mythos-unauthorized-access/
Permanent link: https://aithreat.io/cases/mythos-unauthorized-access/