Raxis is now verified for Red Team Access in Anthropic’s Cyber Verification Program. That gives our engineers Claude Mythos, Anthropic’s newest frontier model, cleared for authorized penetration testing.
The first question we get is whether that means Mythos is doing our pentests. It doesn’t. We use it to build tools. This post explains what that looks like and how our pentesters use it to delve deeper than ever before on our pentests.
What Is Anthropic’s Cyber Verification Program?
Anthropic’s Cyber Verification Program gives vetted security organizations access to Claude models with fewer of the safeguards that block offensive security work. It has three tiers:
- Defense Access for security operations, incident response, malware analysis, and vulnerability validation.
- Red Team Access for everything above, plus authorized penetration testing and red teaming against systems the organization is permitted to test. It is open to organizations only: in-house red teams, government red teams, and penetration testing firms.
- Specialized Access for a small set of organizations authorized to test safety systems, including the members of Anthropic’s earlier Project Glasswing.
Each tier includes Claude Mythos alongside Anthropic’s other current models. Getting Red Team Access means passing Anthropic’s review of the organization, not just the person asking. We did, and that’s part of why we’re writing this.
Why We Build Tools Instead of Running a Model at the Target
Most of a penetration test runs on good, well-known tooling. The hours that matter are where that tooling stops: a proprietary protocol, an appliance with its own management interface, a file format nobody has written a parser for, an authentication flow with a twist. That’s where a real attacker slows down too and where the findings that change a report usually are.
The traditional answer is for the engineer to write a tool on the spot. That works, but it costs hours of the engagement that would rather go to testing. Claude Mythos shortens that. The engineer describes the gap, Mythos drafts the tool, and the engineer reads it, corrects it, and runs it against the lab first.
What we don’t do is point a model at a customer’s environment and let it decide what to attack. A tool built for one exact need does that one thing well, and our Raxis engineer reviews the tool and runs it by hand.
An Example
For example, say an internal network test turns up a building management controller that speaks its own binary protocol over TCP. Commercial scanners see an open port and nothing else.
- The gap. The engineer captures a few sessions between the controller and its management workstation and can see a pattern, but decoding it by hand would take most of a day.
- The tool. The engineer gives Claude Mythos the packet captures as well as the vendor’s public documentation and asks for a parser and a small client that can replay and modify a session. Mythos drafts both.
- The engineer’s part. The engineer reads every line, fixes what Mythos got wrong about the message framing, and tests the client against a captured session in the lab before it touches the customer’s network.
- The test. With a working client, the engineer finds that the controller accepts configuration changes without re-authenticating. That’s the finding. The engineer proves it, documents the evidence, and writes the remediation guidance.
By leveraging Mythos, this task took an hour instead of a day. Raxis customers get a better pentest and a more effective use of engineer time, allowing for lower costs and increased pentest coverage.
Claude Mythos Cybersecurity Work Still Needs a Person
Mythos is very good at writing code for a problem it has been shown. It isn’t accountable for the result. It doesn’t know which hosts are in scope, which test could take a production system down, or which finding the customer’s board will ask about. A Raxis engineer does.
So the rules from the rest of our AI vs Human Pentesting overview still hold:
- The engineer decides what to test. Mythos never chooses targets or runs unattended against a customer network or application.
- Every tool is read before it runs. If the engineer can’t explain what a line does, it doesn’t run.
- Every finding is proven by hand. A Raxis report contains nothing a model asserted without a person confirming it.
Your Data
Red Team Access changes which model our engineers can use. It doesn’t change how we handle customer data:
- All customer-related AI use runs on enterprise accounts under no-training terms.
- Customer data, findings, and reports are never used to train a model.
- Customers who prefer no AI-assisted tooling on their engagement can opt out at scoping, and we honor that for the full term.
Our Trust Center has the full AI use statement.
The Short Version
Raxis has Mythos Red Team Access. We use it to build tools that fit an exact need faster than we could build them by hand. Then a senior engineer runs the test, the same way we always have. If you want to see what that looks like on your environment, talk to us about a penetration test.