Anthropic sets three vetted Claude cyber access tiers

Anthropic merged Glasswing into three Claude tiers that ease cyber refusals for vetted defenders after 129,000 verified flaws.

On October 6, 2026, Anthropic merged Project Glasswing and its Cyber Verification Program into Defense Access, Red Team Access, and Specialized Access, giving vetted users fewer cybersecurity blocks than public Claude models. Outlets agree the lineup includes Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1; Security Affairs and SecurityWeek also name Claude Fable 5.1, and some mention future models. Defense covers incident response, malware analysis, and vulnerability work and is described as open to companies, governments, universities, operators, and researchers; organization-only Red Team Access adds authorized testing but still blocks ransomware and physical-harm or widespread disruption, while Specialized Access, reviewed with the US government, is limited to systems such as power grids, telecom, flight controls, and interbank or payment networks. On CyScenarioBench, Opus 5.5 was blocked on all 50 trials without program access and on 46 of 50 under Defense Access, matching a GBHackers note that Defense completed 4, while Red Team Access completed 34 of 50; The Hacker News and CSO Online say Red Team blocked none, which does not mean every trial succeeded. Glasswing partners reported at least 129,000 verified vulnerabilities from April to July 2026, more than 33,000 high or critical, and Anthropic scanning was described as about 5,500 more findings, with The Register specifying 5,674 true positives (3,014 high and 1,522 critical) of which only 516 were patched; The Decoder said Anthropic claims surveyed impact understates results by at least five times. VulnCheck figures disagree—The Register cited fewer than 0.5 percent of 225 Anthropic-linked flaws exploited in the wild, while The Hacker News said two of 300, CVE-2026-26980 in Ghost CMS and CVE-2026-61500 in Rejetto HTTP File Server—and a later Cyber Security News report said Horizon3 found the Rejetto HFS 3.x flaw with Mythos Preview because predictable session keys can yield forged admin sessions and remote code execution; program access is on the Claude Platform, Vertex AI, and Microsoft Foundry, with limited Amazon Bedrock eligibility and required data retention, and Veracode said about 44% of AI code tasks introduced a vulnerability.