Expanding the Cyber Verification Program
C · Weak signal Vendor product announcement from Anthropic; contains internal evaluation results and partner-reported vulnerability counts that are not independently verified.
Summary based on the full text·Original: about 6 minutes to read
In short
Anthropic announced on 6 October 2026 an expanded Cyber Verification Program that merges Project Glasswing and the earlier program into three access tiers for security professionals. General-availability Claude models keep conservative cyber safeguards that block most offensive work; verified organisations get reduced blocking depending on tier. Anthropic backs the tiering with an internal benchmark run and partner-reported vulnerability counts.
Key numbers
- 129,000verified software vulnerabilities partners reported finding with Claude Mythos, April-July 2026
- 46 of 50CyScenarioBench trials blocked at some point in the Defense Access tier
- 34 of 50CyScenarioBench tasks completed in the Red Team Access tier with no blocks
As reported by the source.
English·Web page·opens anthropic.com
What it reports
- Three tiers: Defense Access (SOC, incident response, malware reverse-engineering; response to applications within a few days), Red Team Access (adds authorised penetration testing; organisations only; review takes a few weeks), Specialized Access (fewest blocks; safety-critical systems such as power grids and interbank infrastructure; reviewed with the US government).
- All tiers include Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1.
- CyScenarioBench, 10 challenges x 5 attempts per tier: without CVP every task blocked on the first prompt; Defense Access blocked 46 of 50 trials; Red Team Access had no blocks and completed 34 of 50, equivalent to the model's 67.6% unrestricted success rate.
- Glasswing partners reported at least 129,000 verified vulnerabilities between April and July 2026; Anthropic's own open-source scanning found 5,500 more between April and October 2026; more than 33,000 rated critical or high.
- Anthropic says the true impact is likely at least five times higher, based on partial data from 33 partner reports; fewer than 50% of partners disclosed patch numbers.
- Data retention is required for monitoring; a zero-retention option (Enterprise Frontier Safeguards) is promised later in the autumn.
How they got it
Internal evaluation of Claude Opus 5.5 on CyScenarioBench with classifiers tuned per tier, plus survey data from a subset of Glasswing partners and Anthropic's own scanning.
Limits
- Vendor-run evaluation and vendor-aggregated partner figures; the "five times higher" multiplier is an estimate.
- Partners triaged differently; patch rates are undercounted.
Why it matters for companies
Security teams will find that general-purpose AI assistants refuse much routine defensive work unless the organisation enrols in a verification scheme with retention and control requirements. Access to the most capable models is becoming tiered by verified identity and use case, which adds procurement and compliance steps. The vulnerability counts, if roughly right, mean third-party software vendors will be shipping a large volume of patches.
Summary based on the full text. Read the original for the full argument and data.Read the original on anthropic.com