Anthropic expands Claude cybersecurity access, adding an authorized red-team testing tier

AT A GLANCE

Anthropic is reportedly set to CVP be changed to three access tiers, allowing approved organizations to conduct authorized penetration and red-team testing.

Article

Anthropic Anthropic Claude opens “Red Team Access”: security organizations can conduct authorized penetration testing

Beating AI news flash,Anthropic expands Cyber Verification Program（CVP), merging the original single-tier access and Project Glasswing into three tiers. Security teams that pass review can obtain fewer cybersecurity safeguards on Claude Opus 5.5、Sonnet 5.5 and Mythos 5.1. The newly added Red Team Access explicitly permits authorized penetration testing and red-team testing.

The three tiers are Defense, Red Team, and Specialized. Defense is intended for incident response, malware reverse engineering, and vulnerability analysis, and individual security researchers may also apply. Red Team is open only to organizations, which may test authorized systems. Specialized has the fewest restrictions and is primarily for strictly vetted organizations testing critical systems such as power grids, aviation systems, telecommunications networks, and bank transfers.Anthropic used Opus 5.5 to run 50 complex cyberattack tasks. The standard version was blocked in the first round every time across 50 attempts, while the Defense tier was still blocked partway through in 46 attempts; in the Red Team tier, none of the 50 attempts triggered a security safeguard, ultimately completing 34 attempts, with a success rate broadly equivalent to when these safety restrictions were fully disabled. However, operations such as ransomware and physical-system sabotage that could cause serious harm remain prohibited.Anthropic previously made Mythos’s advanced cybersecurity capabilities available to only a small number of organizations through Project Glasswing. Partners April to July confirmed at least 129000 software vulnerabilities, more than 33000 of which were high-severity or critical vulnerabilities. Glasswing is now incorporated into CVP, and existing members are directly placed in the highest Specialized tier.

Original link https://m.theblockbeats.info/flash/370556

Key points

Anthropic expands Cyber Verification Program, merging the original single-tier access and Project Glasswing into three tiers: Defense, Red Team, and Specialized.
Security teams that pass review can obtain fewer cybersecurity safeguards on Claude Opus 5.5、Sonnet 5.5 and Mythos 5.1.
Red Team Access It allows organizations, but not individuals, to test authorized systems.
Defense is intended for incident response, malware reverse engineering, and vulnerability analysis, and individual security researchers may also apply.
Specialized has the fewest restrictions and is primarily for strictly vetted organizations testing critical systems such as power grids, aviation systems, telecommunications networks, and bank transfers.
The article says that Anthropic used Opus 5.5 to conduct 50 tests of complex cyberattack tasks: the standard version 50 was blocked in the first round every time, the Defense tier was blocked partway through in 46 attempts, while the Red Team tier triggered no security safeguards and completed 34 attempts.
Operations such as ransomware and physical-system sabotage that could cause serious harm remain prohibited.
The article says that Project Glasswing partners, from April to July confirmed at least 129000 software vulnerabilities, more than 33000 of which were high-severity or critical vulnerabilities.
After Project Glasswing was incorporated into CVP, existing members were directly placed in the Specialized tier.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This adjustment does not remove all safety restrictions; instead, it provides different levels of cybersecurity access based on the applicant and intended use. Approved individuals or organizations can use the relevant models for vulnerability analysis, incident response, or authorized security testing according to their tier. Operations that could cause serious harm remain prohibited.

Why it matters to readers

Fewer security safeguards could allow approved security teams to perform more comprehensive testing within the authorized scope. The test results in the article show that the Red Team tier completed complex cyberattack tasks more often than the standard version and the Defense tier, but these data have not yet been verified against primary materials.

Beginners should clearly distinguish “authorized security testing” from unauthorized attacks: the Red Team tier allows organizations to test only systems for which they have authorization, and fewer safeguards do not mean they may attack targets indiscriminately.

Risks and unknowns

The evidence comes from a media news flash and does not include Anthropic an official announcement or specific application documents.
The model names, three-tier access rules, and test results cannot be independently verified based on the current evidence.
50 The specific content of this task, its evaluation criteria, and the definition of “completion” have not been disclosed.

Anthropic expands Claude cybersecurity access, adding an authorized red-team testing tier

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center