Anthropic Enhances Cyber Verification Program, Expanding AI Access for Security Teams

Anthropic has made significant strides in its Cyber Verification Program, broadening access to its advanced AI models for security teams with a reduced set of restrictions. This initiative is poised to empower cybersecurity professionals by allowing them enhanced capabilities to defend systems more effectively.

The expanded program introduces a tiered access system based on the specific cybersecurity roles that organizations hold. There are three distinct tiers: Defensive Access, Red Team Access, and Specialized Access. Each tier aligns with the organization's authorized activities, ranging from defensive security measures to high-stakes testing for critical infrastructure.

“Defenders also need access to the best tools and most powerful capabilities to secure their systems,” Anthropic articulated in its announcement, highlighting the necessity of advanced tools in today’s cybersecurity arena.

Currently, Anthropic's models come equipped with cautious cyber safeguards that are designed to thwart a majority of potential attacks. These restrictions aim to minimize harmful usage while Anthropic continues to work on reducing false positives that can disrupt secure coding initiatives.

Bridging Initiatives: From Glasswing to CVP

The Cyber Verification Program is the result of amalgamating two previous initiatives—Project Glasswing and the earlier CVP. Project Glasswing provided select organizations engaged in securing vital software access to Claude Mythos, while the CVP offered access to Claude Opus and Claude Sonnet models with limited safeguards.

In its new structure, existing members from Glasswing will transition to the Specialized Access tier without needing to reapply for their current models. This move is part of Anthropic’s effort to increase the influence of Project Glasswing while engaging a broader community of cyber defenders.

The Defense Access tier focuses on activities such as security operations, incident response, malware reverse engineering, and vulnerability testing. The Red Team Access tier further enables users to engage in authorized penetration tests with predefined scopes. In contrast, Specialized Access holds the least restrictions and is designated for verified organizations conducting tests on critical systems, including those related to aviation, power, telecommunications, finance, and government.

Notably, that access is granted under stringent scrutiny and often in collaboration with government bodies.

Evaluating the Access Tiers

Anthropic assessed the efficacy of different access levels through a structured evaluation called CyScenarioBench. This benchmarking tests the ability of AI models to execute complex cyber operations over several stages.

In these tests, Claude Opus 5.5 faced a series of ten challenges with five attempts for each tier. Remarkably, without CVP access, every single trial was blocked. Under Defense Access, 46 out of 50 trials faced some level of blockage. However, there were four successful completions.

In contrast, the Red Team Access status allowed all attempts to go through successfully, with Claude completing 34 challenges—reflecting an equivalent success rate of 67.6% absent any safety safeguards. Nevertheless, caution is advised in interpreting these results, with IDC's research director Sakshi Grover emphasizing the importance of context, as the trials were vendor-conducted.

Grover further pointed out that proper authorization for the use of AI in security testing must be clearly defined to differentiate between legitimate uses and potential misuse, considering the close resemblance of these processes to those employed by malicious actors.

Enhancing Control Measures

Further reinforcing the conversation, Grover advocated for the implementation of controls beyond the realm of the AI model itself when decreased safeguards are in play. She suggested that each AI agent should carry a unique identity, equipped with temporary permissions aligned exclusively with their specific tasks. Evaluating actions and ensuring containment capabilities are key to maintaining security.

Deepika Giri, an Asia/Pacific AI advisory VP at IDC, corroborated this outlook, recommending enterprises treat AI agents similarly to privileged staff members, implementing full oversight for significant actions to mitigate risks.

Cybersecurity expert Vibhum Dubey added that clarity around accountability for actions conducted by autonomous systems is imperative for organizations leveraging these advanced tools.

Data from Project Glasswing indicated that partners uncovered at least 129,000 verified software vulnerabilities between April and July 2026, with Anthropic's own investigations revealing an additional 5,500 vulnerabilities in the same timeframe. Of these, more than 33,000 were designated critical or high severity, although these figures may significantly underrepresent the true count due to their basis on survey data from just a subset of partners.

Organizations participating in the CVP will need to comply with data retention policies as Anthropic seeks to monitor for misuse. Furthermore, the forthcoming Enterprise Frontier Safeguards service is expected to enable eligible organizations to integrate zero-data-retention measures with additional protections, all while allowing them to maintain control over their data in cloud environments.