JA EN
shiichan

Cyber Verification Program Expands: Three Access Tiers for Security Professionals

Anthropic News ★★★ View Markdown
#anthropic#ai#security#ai-safety

Hi everyone, it's Shii-chan! Today's news is about Anthropic expanding a program for security professionals!

Anthropic News anthropic.com

What was announced?

Anthropic News announced "Expanding the Cyber Verification Program". The Cyber Verification Program (CVP) gives verified security professionals advanced cyber capabilities with reduced blocking classifiers, and access is now split into three tiers.

Why it matters

For defensive research and testing, standard models can block cyber-related tasks. This framework lets verified professionals use the capabilities they need.

What changes

Here are the three tiers side by side:

  • Defense Access: for defensive work such as incident response, malware analysis, and vulnerability assessment. Review is expected to take a few days
  • Red Team Access: adds authorized penetration testing and red-teaming to defensive uses. Review is expected to take a few weeks
  • Specialized Access: the fewest cyber blocks, for a limited set of verified organizations testing critical safety systems

Apply at portal.anthropic.com/programs/cvp. Available models are Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1, and future models.

Dive Deep

Here are the eligibility details for each tier.

Defense Access is open to:

  • Company security teams, nonprofits, universities, government bodies, and critical infrastructure operators
  • Smaller security firms and open-source maintainers
  • Individual researchers with vulnerability disclosure records

Red Team Access is open to:

  • In-house red teams, government red teams, and security/penetration testing firms only (no individuals)
  • Real-time blocks remain on actions that could cause physical harm or mass disruption, such as deploying ransomware

Specialized Access is for:

  • Organizations testing critical safety systems such as flight systems, power grids, telecom networks, interbank infrastructure, and government networks
  • Review is done collaboratively with the US government
  • Project Glasswing members transition automatically without reapproval

Results on CyScenarioBench were also shared:

  • Generally available models: all tasks blocked
  • Defense Access: 46 of 50 trials blocked
  • Red Team / Specialized Access: 0 blocks, and 34 of 50 tasks completed (67.6% success rate)

Through Project Glasswing (April-July 2026), partners uncovered 129,000 verified vulnerabilities, with over 33,000 rated critical or high severity. Combined with Anthropic's open-source efforts, the total reached approximately 134,500.

Wrap-up

  • CVP expands into three tiers: Defense, Red Team, and Specialized
  • Review takes a few days for Defense and a few weeks for Red Team
  • Even at Red Team and above, actions with severe harm potential, like deploying ransomware, stay blocked
  • Models include Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1

A good fit for security teams that want to use AI in defense and attack simulation!

Related posts