Secure.com Joins Anthropic’s Cyber Verification Program

TL;DR

Secure.com has been accepted into Anthropic’s Cyber Verification Program (CVP), which gives our security research team verified access to Claude’s dual-use capabilities (the offensive reasoning that frontier models gate by default) for legitimate defensive work.

We are announcing two things at once: our acceptance into CVP, and the CVP Landscape Report 2026, our independent map of how publicly announced companies are using verified frontier capability, what they are using it for, and what it means for the industry and their customers.

The program lets vetted cybersecurity organizations use Claude for high-risk, dual-use activities such as vulnerability research, penetration testing, exploitability analysis, and authorized red teaming. Prohibited activities, including ransomware development and mass data exfiltration, stay blocked for everyone.

For us, this is not just a capable model, but it sharpens the intelligence behind Secure.com’s Red Teammate and the offense-to-defense loop that runs across the platform. Every validated offensive finding becomes shared defensive context that our SOC, Cloud Security, AppSec, Risk, and Compliance AI Teammates can act on, which prioritize, assign, remediate under governance, evidence, and retest.

Being vetted by the team that built the model is an independent verification of how we operate, instead of a marketing badge.

What CVP Actually Is, and What It Isn’t

Anthropic applies real-time safeguards to its most capable models. Some security work sits in a gray zone: the same reasoning that helps a defender model an attack path also helps an attacker build one.

Anthropic sorts that work into two buckets.

• Prohibited use (ransomware development, mass data theft) stays blocked for everyone, verified or not.
• High-risk dual use (vulnerability research, exploitability analysis, adversarial simulation) can trigger those safeguards by default, but legitimate defenders can apply to adjust them.

CVP is that application. It is free, organization-scoped, and reviewed by Anthropic. Approval ties to one specific organization and clears only dual-use work; the prohibited categories never open up.

Why We Wanted In

To defend a system, you have to understand how it gets broken, and how a real attacker finds a flaw, reaches it, and chains it into a breach.

A scanner can tell you a vulnerability exists but it can’t tell you whether:

• It’s reachable
• The surrounding configuration makes it exploitable
• An identity weakness enables escalation
• Several low-severity findings combine into one real path
• Your SOC would even see the resulting activity.

Without that reasoning, prioritization is an informed guess, teams patch what looks severe instead of what creates the clearest path to impact. Secure.com fixes issues, not just flags them. Fixing them well means understanding them the way the attacker does. We applied for CVP because the work needs that depth; not because the program happened to exist.

The Loop in Practice: Predict, Prove, Govern, Retest

The access sits at our internal research layer. It sharpens the intelligence that feeds Secure.com’s Red Teammate, which operates as a governed offensive teammate against authorized targets. The product vision is four steps:

• Predict the path. Use shared asset, identity, vulnerability, application, configuration, and business context to identify the routes an attacker is most likely to take.
• Prove what is real. Test authorized paths and separate reproducible exploitability from theoretical exposure, scanner noise, and disconnected severity scores.
• Govern the fix. Route the result to the right defensive AI Teammate, owner, workflow, SLA, control, and approval step, then retest to confirm the exposure is closed.
• Retest continuously. After remediation, run the same path again, so posture improves through every exercise instead of decaying between annual assessments.

CVP participation gives our research and product teams more room to study the hard middle step: how sophisticated attackers reason across weaknesses and turn individual findings into a coherent chain.

How Verified Reasoning Sharpens the Platform

• Deeper exploitability analysis: reasoning through whether a disclosed flaw is actually exploitable in a real environment, which is what separates a noisy finding from a genuine risk.
• Sharper attack-path modeling: mapping how an exposed asset, a KEV vulnerability, and a misconfigured IAM role could reach a crown-jewel system, and estimating blast radius.
• Stronger remediation logic: better inputs produce better fixes, and the gap between what a scanner flags and what an attacker exploits is exactly where teams get hurt.
• Measurable defensive coverage: every authorized Red Teammate action maps to MITRE ATT&CK and correlates with the customer’s detection telemetry, so teams can see what was detected, what was only partially caught, what passed unnoticed, and where rules or telemetry are missing.

And because validated findings connect to the same shared context, the work doesn’t stop at a report:

• SOC Teammate investigates the activity, improves detection logic, and coordinates response.
• Cloud Security Teammate addresses cloud posture, configuration, identity, and network weaknesses.
• AppSec Teammate traces runtime exposure back to vulnerable code, dependencies, containers, or IaC.
• GRC Teammate maps the exposure and remediation evidence to applicable controls, exceptions, and audit requirements.

Greater capability is paired with tighter control. Offensive testing stays scoped to authorized environments; high-impact defensive actions stay explainable, reviewable, auditable, and subject to human approval where required. The point of the AI is to accelerate investigation, validation, prioritization, and execution, not to remove accountability.

What This Signal Is Actually Worth

Anthropic has framed its goal as building a durable advantage for defenders over attackers, and we share it. Threat actors already lean on open and uncensored models to speed up their work; what they can’t reach is verified access to frontier models. That gap is where a real defender advantage lives.

So treat CVP verification as a useful but specific signal. It confirms that the research behind the product cleared an independent review by the people who built the model (a different bar than a claim about being “AI-powered”). It vets our dual-use access. It does not, on its own, prove one product beats another, so weigh it alongside everything else you measure.

The Bigger Shift Underway

Frontier AI is making vulnerability discovery and adversarial reasoning faster, and Anthropic has argued that defenders will need to adopt AI with comparable sophistication and shorten the time between finding a weakness and fixing it. Access to more capable models is only part of the answer. The harder part is turning that capability into an operating model organizations can safely use:

• Authorized rather than open-ended.
• Continuous rather than annual.
• Evidence-based rather than severity-based.
• Connected to defense rather than trapped in a red-team report.
• Governed rather than uncontrolled.
• Measured by closed exposure rather than the count of findings produced.

That is the contribution we intend to make. We are not running offense as a separate service that ends when the assessment does. We are building offense into the start of a continuous defensive cycle.

• Every attack path becomes a defensive context.

• Every validated exploit becomes prioritized work.

• Every remediation becomes evidence.

• Every retest makes the next attack harder.

We Mapped the Field: The CVP Landscape Report 2026

We wanted to see the whole board, not just our own square. So alongside our acceptance, we are publishing the CVP Landscape Report 2026, the first public map of how the security industry is putting verified frontier cyber capability to work.

CVP has no official public roster, which has made the picture hard to see. Our research team independently identified the companies that have publicly announced acceptance, then coded each announcement by security category, declared use case, customer outcome, and governance posture. The result is the clearest available view of an emerging capability layer.

We are explicit about what the report is and is not. It is a publicly observable sample, not an official Anthropic participant list, so category concentration may reflect who chose to announce rather than the full CVP population. What it captures is a direction of travel: verified frontier capability is becoming a defensive layer, and the organizations adopting it are increasingly wiring it into how work actually gets closed.

For customers evaluating vendors, the report offers a neutral frame for what verified access does and does not signal. For the industry, it is a shared starting point for a conversation that has so far happened one announcement at a time.

Read the CVP Landscape Report 2026

Governed Defense Powered by Offense

Secure.com is an AI-native security operating layer above the stack you already own. Its AI Teammates share context across assets, identities, vulnerabilities, cases, risks, controls, attack paths, and workflows, so offensive findings become governed defensive action across the security lifecycle. 

CVP strengthens our ability to pursue that mission: not by giving AI unlimited authority, but by giving verified defenders more of the reasoning they need, surrounded by the context, governance, human oversight, and execution required to turn intelligence into protection.

Attackers are using AI to connect the kill chain. Secure.com is using AI to close the loop.