BreachLock Joins Anthropic’s Cyber Verification Program (CVP)

Anthropic’s Claude models, as they stand today, actively block offensive security work by default. While it’s intended to be utilized for legitimate purposes by cybersecurity professionals like pentesters to find vulnerabilities, it could also help an attacker exploit those same vulnerabilities if it ended up in the wrong hands. The model can’t differentiate between legitimate and illegitimate use cases based on prompts alone, so Anthropic built a classifier that blocks these capabilities for everyone by default and created an exclusive program to verify the organizations that actually need it.

That program is the Cyber Verification Program (CVP), and BreachLock has been accepted into the CVP as of July 2026.

What is the Anthropic CVP?

CVP is Anthropic’s application-based process for vetting security organizations. When Anthropic rolled out real-time cyber safeguards on its Claude models, it established two categories of blocked activity:

1. Prohibited use: Activities with essentially no legitimate defensive application, like ransomware development, command-and-control infrastructure, and mass data exfiltration. These currently remain blocked for everyone, including CVP members.

2. High-risk dual-use activity: Work that has clear legitimate defensive applications but also overlaps with offensive techniques, like vulnerability exploitation analysis, adversarial simulation, threat modeling, and red team workflows. These are also blocked by default, but CVP exists to lift that restriction for approved organizations, like BreachLock, that can demonstrate a legitimate need.

To get approved for the CVP, organizations must apply by describing their organization and specific use case in detail. Anthropic then reviews its legitimacy, and if approved, the restrictions are adjusted on a case-by-case basis for specific organization ID, scoped to the use case described, and subject to Anthropic’s usage policy.

Why the Anthropic CVP Exists

Most security leaders have already seen the Glasswing headlines. Anthropic’s Claude Mythos Preview model, released in April of 2026 under controlled access, surfaced over 23,000 issues across more than 1,000 open-source projects, with over 6,000 of those rated high or critical. 90%+ of the vulnerabilities identified were independently validated to be true positives, which solidified its legitimacy.

The model proved it could find vulnerabilities at scale and, more importantly, understand how to exploit them. That is why Anthropic chose to restrict the capability by default and build a verification process for organizations that need it for defensive work.

That verification process is now known as the CVP. It’s something worth paying attention to even for those who don’t use Claude, because it sets a precedent for the even more powerful capabilities that are yet to come in the future. Anthropic is the first major AI provider to build a formal verification process for security practitioners. For anyone managing an offensive security program, that precedent is worth watching.

What Joining Anthropic’s CVP Means for BreachLock

BreachLock operates across penetration testing services, adversarial exposure validation, red teaming, and attack surface management. All of that work requires reasoning about how attackers actually operate, including but not limited to how they chain vulnerabilities together, how they escalate privileges, how they move laterally through environments, and how they exfiltrate data.

While BreachLock has spent over seven years building and deploying its own offensive security technologies in production environments, including autonomous penetration testing, that same category of reasoning is exactly what Anthropic’s classifiers block by default when applied through Claude. For a company whose entire product line is built on offensive security methodology, those default blocks create friction in research workflows, tool development, and the kind of adversarial simulation that makes our testing more effective.

BreachLock’s acceptance into CVP removes that friction for our approved use cases. Our security research team can now use Claude’s dual-use capabilities for exploitation analysis, adversarial simulation, and offensive tooling research without hitting the default safeguards that restrict that work for general users. In practice, that means fewer interruptions when our researchers are working through attack chain logic, building proof-of-concept exploits for client engagements, or testing how an adversary would move through a specific environment. The access is scoped to our organization, monitored against our stated use case, and governed by Anthropic’s usage policy.

This does not change what we do. It makes the capabilities we have built over the last seven years even stronger.

What this Signals for Security Leaders

The CVP is worth thinking about beyond BreachLock’s own acceptance into it, as AI models are getting better at security reasoning day by day.

For security leaders evaluating how AI fits into their programs, CVP is an early example of what the vendor relationship is going to look like going forward. Access to the most useful capabilities will increasingly require organizations to demonstrate who they are, what they do, and how they plan to use the tools. This will inherently come with implications for procurement, for compliance documentation, and for how security teams think about their AI toolchain.

It also becomes a signal about vendors themselves. If an AI provider is reviewing organizations before granting access to its most capable security features, then acceptance into that process says something about the vendor being reviewed, not just the tools they get access to.

As AI models become more capable at security reasoning, the gap between what a verified organization can do with those tools and what an unverified one can do is only going to widen. For offensive security practitioners specifically, getting verified now, while these programs are still relatively new, is a practical step towards narrowing that gap.

Organizations that invest early in building trusted relationships with AI providers will have a real advantage as these capabilities continue to mature.

Conclusion

BreachLock’s acceptance into the CVP is a practical improvement to our research and development workflow. It gives our team better tools for the work we already do and also reflects a broader shift in how the industry is starting to manage the intersection of AI capability and security practice. The organizations that will benefit most from this shift are the ones that can demonstrate, under review, that their work is legitimate, scoped, and responsible.

That has always been how BreachLock operates, and CVP is a clear, independent validation of that.

Author

BreachLock Labs