Claude Mythos access is restricted. That does not mean verified users receive a model with no safeguards. This owner now reflects Anthropic’s October 2026 Cyber Verification Program (CVP), replacing earlier blanket descriptions of Mythos as “unrestricted” and claims that only government-connected organizations could apply.

What changed in October?

Anthropic’s October 6 announcement expands CVP to three access tiers covering Opus 5.5, Sonnet 5.5 and Mythos 5.1, with future models also subject to evaluation. Reduced blocking for verified cyber work is permissioned access, not general availability or immunity from the usage policy.

Tier	Intended work	Eligibility boundary
Defense Access	Defensive security, including incident response and vulnerability analysis	Qualifying organizations; verified independent researchers can apply
Red Team Access	Authorized penetration testing and exploit validation	Organizations only
Specialized Access	More sensitive, approved security work	Closely vetted organizations; additional review with the US government

Existing Glasswing and CVP members do not need to reapply for their current access. Transition and provisioning for newer models must still follow the program instructions. Higher access tiers retain restrictions, including safeguards against physical harm and mass disruption.

How to apply and provision access

The CVP help article directs applicants to Anthropic’s Verification Portal: open Programs, choose Cyber Verification Program and apply. Organizations apply once; administrators assign access rather than having every employee apply separately. Individuals are eligible only for Defense Access, not Red Team or Specialized Access.

For API organizations, an owner must assign the approved program to the appropriate workspace. Cloud-provider access requires linking the relevant account; availability and provisioning can differ by platform. A consumer subscription or ordinary API key alone does not establish a CVP grant.

Generally available Claude models still support secure code review, threat modeling, patching and analysis of your own source. Apply when legitimate professional security work is blocked, not merely to obtain a different model label.

Security and data handling

Read the tier-specific security requirements before integrating an approved grant. They cover authentication, credentials, approved users, access control and incident handling. Requirements rise with the access level; a personal API-key workflow should not be assumed compliant with an organizational grant.

CVP normally requires retention for misuse monitoring. The help article documents a temporary exception for organizations with existing Fable or Mythos zero-retention access, pending Enterprise Frontier Safeguards. Do not assume a standard zero-data-retention arrangement automatically extends to every new grant. Enterprise Frontier Safeguards is described as forthcoming, not something every account can enable today.

Practical governance checklist

• Confirm the approved tier, user and workspace before investigating apparent blocks.
• Restrict activity to owned or explicitly authorized systems.
• Verify retention, credential and cloud-provider requirements with the program documentation.
• Keep an audit trail of permissions and do not expose model credentials to unapproved users.
• Separate approved research access from any client-facing productization permission.

For the general development context, see our Fable model guide and AI agent security guide. Neither grants access to CVP.

Verified October 7, 2026. This is an access/governance reference, not a hands-on evaluation or a guarantee of approval.