Expanding the Cyber Verification Program

What it is

Anthropic's expanded Cyber Verification Program (CVP), announced in 2026, replaces Project Glasswing and the earlier CVP with one program. Every tier covers Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1 and later models. Generally available models (Opus 5.5, Fable 5.1, Sonnet 5.5) keep conservative cyber safeguards that block most cyber work.

Which tier fits

Red Team applicants are enrolled in Defense Access while under review. Red Team users still hit real-time blocks on actions that could cause physical harm or mass disruption, such as deploying ransomware, damaging physical systems, or pen testing high-risk safety systems. Existing Glasswing members move to Specialized Access without reapproval for current models.

Generally available models still handle code review, patching known issues, vulnerability finding in owned source code and alert triage, so many teams may not need CVP for that work.

Data retention and platforms

Enrolled organizations must allow data retention so Anthropic can monitor for misuse. Enterprise Frontier Safeguards (EFS), due later in fall 2026, will let eligible organizations keep data in cloud infrastructure they control. Until then, organizations with zero data retention access to Fable 5.1 or Mythos 5.1 can use CVP with zero data retention.

CVP is available on the Claude Platform, Google Cloud Vertex AI and Microsoft Foundry. On Amazon Bedrock it is offered only to customers eligible for EFS.

What to do

• Apply through the program form. Anthropic verifies all applicants and asks for proof of the security controls required for the tier.
• Existing CVP members keep current settings for earlier models and are evaluated automatically for Opus 5.5, Sonnet 5.5 and Mythos 5.1.
• Admins must assign access to specific workspaces.
• Teams that need zero retention or Bedrock should register interest in EFS.
• Report cases where a block hits work the tier should allow, using the page's reporting form.

Evaluation results

Anthropic ran Opus 5.5 on CyScenarioBench, a test of multi-stage cyber operations, with 10 challenges and 5 attempts each per tier (50 trials).

flowchart LR
  A[No CVP] --> A1[All blocked on first prompt]
  B[Defense Access] --> B1[46 of 50 blocked, 4 succeeded]
  C[Red Team Access] --> C1[No blocks, 34 of 50 completed]
  D[No safeguards] --> D1[67.6% success]

Red Team's 34 of 50 completions is 68%, which Anthropic calls effectively equal to the 67.6% rate with no safeguards, taken as representative of Specialized Access. Anthropic says the pattern was the one it expected.

Caveats

• The test covers one model (Opus 5.5) on one benchmark, run by Anthropic. It measures blocking of offensive scenarios, not false positives on legitimate defensive work.
• The 129,000 verified vulnerabilities (April to July 2026) come from survey data of a subset of Glasswing partners. Anthropic's own open-source scanning added 5,500 between April and October 2026. More than 33,000 were rated critical or high severity. Anthropic expects the true impact to be at least five times higher.
• Tier requirements and controls are summarized only briefly. Full details are in Anthropic's Help Center.