CISO Voice

Exclusive / News

Netskope CISO James Robinson on Turning an Unreleased Anthropic Model Loose on His Own Code 

We speak to a CISO inside Project Glasswing on what Claude Mythos found in code his own team had already tested, and why defenders still have the advantage.

Mirren McDade
Mirren McDade Senior Feature Writer · November 17, 2026 · 6 min read
Netskope CISO James Robinson on Turning an Unreleased Anthropic Model Loose on His Own Code

“One early principle I was taught in pen testing and hacking is that everything can be broken into. You just need enough time, enough focus and attention, and some skill,” says James Robinson, CISO at Netskope. 

Robinson has spent more than 25 years in security, starting out in computer networking and penetration testing before moving into product security at Emerson Electric, then on to Websense and Optiv. He joined Netskope as Deputy CISO in October of 2018, and has now held the top job for around two and a half years. 

Netskope is one of the organizations lucky enough to be taking part in Project Glasswing, Anthropic’s invitation-only program for companies that build critical software. It gives them early access to Claude Mythos, a model Anthropic has yet to release. 

Anthropic claims their model can outperform most humans at finding vulnerabilities. Robinson led the effort to test that claim himself by pointing the model at Netskope’s own products. 

“For me, it was validation that these capabilities are real. I went into it a little skeptical, thinking, ‘Really? Let’s test it and see.'” 

Netskope’s products had already been tested by people who knew how to break things. The model still managed to uncover nine kernel-level bugs in them, and seventeen more in the open-source code underneath. 

Letting the model choose its own way in 

Robinson’s team gave the model a target inside a trusted environment, something along the lines of “break into our web UI,” and left it up to the model which route to take. It could choose to come through the APIs, the web UI, or the CLI when they pointed it at one of Netskope’s appliances. At one point it started trying to reach the appliance over other protocols entirely. The team went back and examined those afterward. 

That freedom was the point. Testing the same places they always tested would turn up the same things they always found, so they wanted the model working from its own analysis rather than from a script. 

They set firm limits that there would be no production testing, and no source code at first. The model started out with design documents and similar material, then was granted more access inside the test environment as the exercise went on. “We were treating the model like a pen tester, stepping it up from black box toward white box testing and giving it more and more visibility.” 

What came back 

Pointed at two of Netskope’s shipping Windows kernel drivers, the model produced nine kernel memory corruption findings. It also found 17 previously unreported vulnerabilities in the open-source C and C++ libraries the Netskope product is built on. 

“We’d been building up team members to do advanced fuzz testing and security testing,” he says. “When we paired those team members with a more than capable model, like Mythos or OpenAI’s cyber model, they started finding more advanced things we just hadn’t been able to get to: multi-step attacks, and RCEs we hadn’t previously discovered. That was primarily down to time, not capability. It was the speed at which the model could work through it.” 

“But the model couldn’t do it alone,” Robinson adds. “I don’t think we’d have gotten there just by saying, ‘Go break into my tooling.’ What made it work was plugging the model into the fuzz testing frameworks and harnesses we’d spent years building, then letting it branch out into other areas from there.” 

“The surprising part was who was finding these issues,” he says. “These were people with five or seven years of experience, not 15 or 20, and they were uncovering much more advanced bugs. That was an eye-opening moment.” 

 When the fix isn’t yours to make 

Seventeen of the findings were not actually in Netskope’s own code. They were in open-source projects its products are built on, maintained by people working without deadlines or budgets. This meant scanning for the issues may have been faster, but fixing them had some hold ups. 

That creates a gap Robinson cannot close from inside his own company, with cost being a big factor. Netskope is exploring partnerships with a US government project that could help get those patches funded and maintained, but nothing yet solves the underlying problem: nobody owns these gaps. 

“The industry needs to come together around some of these open-source projects.” He points to Google’s Project Zero, which hunted flaws in core internet infrastructure, and says he has given both Anthropic and OpenAI the same guidance: build that kind of community around what these models are now turning up. 

Until that exists, Netskope will continue taking its findings to the open-source community through its threat labs. Where staff are already active in those communities, it has sometimes fixed the flaws itself, though Robinson notes that is a luxury he did not have at previous employers. 

For everyone without an invitation 

For CISOs outside the program, Robinson’s first bit of advice is to check the partners they work with, as a contractor or consultant may have access already.  

After that, he wants to see teams getting hands-on with what they can obtain. Netskope has tested open-weight models such as Kimi and GLM on compute it built itself, alongside OpenAI’s cyber model and now Google’s. “If you can host your own models and test with open-weight models, do it. Let your team go and play. Build a GPU stack with five or 10 GPUs and test that way.” 

He expects the access gap to close on its own over time. By his reckoning the industry is one or two hardware generations away from running these models, or some version of them, on ordinary machines, at which point keeping them off devices becomes the same losing battle as keeping penetration testing tools off laptops. 

That is the shift he wants other CISOs to start preparing for. In his mind, APT no longer only means “advanced persistent threat”. It also means “agentic persistent threat”, due to how AI agents are constantly trying to break into organizations. 

Robinson thinks defenders still hold the advantage, as long as they can maintain visibility. Netskope tests that by breaking real attacks into steps and threat modeling them against its own environment, using the Cyber Kill Chain as a baseline. The AI-driven attack on Hugging Face ran to 17,000 actions over four and a half days, with one zero-day and three data sets exfiltrated. Robinson has no doubt his SOC would spot a single entity doing that, because it would stand out on any chart. 

“As it stands today, and this could change very fast, these models are very chatty and aren’t trying to be covert. So as long as you have the visibility, you have a strategic advantage against some of these attacks.” 

Mirren McDade
Mirren McDade Senior Feature Writer

Mirren is an experienced reporter with a background in a range of industries, including cloud business technologies, cloud security, information security and cyber security, and has conducted interviews with several industry experts. 

More from CISO Voice

Sign up to CISO Voice, the free newsletter trusted by senior security leaders.

Secret Link