Skip to content

Anthropic

Anthropic splits Mythos access into three security tiers

3 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

On October 6, Anthropic merged its two security-access programs into a revamped Cyber Verification Program, Reuters reported. The new program has three tiers, and every tier gives vetted teams Claude Opus 5.5, Sonnet 5.5, Mythos 5.1 and future models with fewer safeguards. If your team does incident response or malware analysis, the Defense tier is the one to apply for, and researchers with a record of reported vulnerabilities qualify too. Penetration testing and red-teaming sit in a separate tier that only organizations can apply for.

The program replaces two that Anthropic ran for the past six months. Project Glasswing gave organizations securing critical software access to Claude Mythos, Anthropic’s most cyber-capable model family. The original CVP gave vetted security teams reduced safeguards on Claude Opus and Sonnet. When Anthropic unveiled Claude Mythos Preview in April, it raised fears that AI could hack software before that software had been secured.

Who can apply for what

Each tier has its own verification requirements and security controls. Defense covers work such as incident response and malware analysis. It is open to security teams, critical infrastructure operators, open-source maintainers and researchers with a record of reported vulnerabilities. Anthropic aims to answer Defense applications within a few days.

Red Team adds authorized penetration testing and red-teaming. Its reviews may take several weeks, and applicants get Defense access while they wait.

Specialized has the fewest restrictions. It is reserved for a small group of organizations authorized to test safety-critical systems such as power grids, flight systems and interbank transfer infrastructure, and Anthropic vets each member together with the US government. Existing Glasswing members move into this tier without reapproval for current models.

Even at the top tier, real-time controls still block ransomware deployment, damage to physical systems and pentesting of high-risk safety systems.

The 129,000 count

Glasswing’s numbers came before the expansion. Partners found at least 129,000 verified vulnerabilities between April and July, and Anthropic’s own open-source scanning found 5,500 more between April and October. More than 33,000 have been rated critical or high severity so far. Anthropic says the figures are likely an undercount and expects the true impact to be at least five times higher. It also notes that the data comes from a survey of a limited number of partners.

Anthropic has also described the other end of the pipeline. According to CSO Online, the company said some open-source maintainers are severely capacity constrained and have asked it to slow down disclosures. Anthropic also wrote that finding vulnerabilities is relatively easy compared with fixing them. Over six months, Anthropic’s models flagged more than 29,000 candidate vulnerabilities in major open-source projects, and about 6,000 of them got human review. A disclosure dashboard cited by The New Stack showed 516 patched upstream as of October 2.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.