The Engineers Building Both Sides of the AI Security War
A new breed of security team is quietly writing the rules for how artificial intelligence gets used in cyberattacks and defenses. Most companies have never heard of them.
Key points
- Anthropic invited more than 50 organisations in April to test Claude Mythos, its advanced AI security tool, under a programme called Project Glasswing.
- OpenAI launched a parallel programme called Daybreak, giving selected organisations early access to GPT-5.5 for security testing.
- Cisco has open-sourced its AI security harness, called Foundry Security Spec, built around 13 specialised AI agents and roughly 130 individual requirements.
- Security teams using these tools have already uncovered thousands of vulnerabilities, including some flaws undetected in core technologies for decades.
- "Yellow teams", a term coined at Black Hat 2017, are the engineering groups now responsible for turning AI security tools into something usable in production.
Two of the most powerful AI security tools ever built are currently in the hands of a small number of companies. The engineers inside those companies are learning something uncomfortable: the same technology that finds your weaknesses can be handed to someone who wants to exploit them.
Anthropics Claude Mythos and OpenAIs GPT-5.5 aren't consumer chatbots. They're purpose-built AI systems designed to hunt for software vulnerabilities at a scale no human team can match. Early access came through Project Glasswing and the Daybreak programme respectively. The organisations inside those programmes are still figuring out what they're holding.
Why does this matter to people who aren't security engineers?
Because the same tools finding holes in company systems right now will eventually reach criminals. Every bank, hospital, and retailer that runs software sits downstream of whatever these teams figure out in the next twelve months.
In practice, the AI doesn't arrive ready to use. James Robinson, the chief information security officer (the senior executive responsible for a company's security) at Netskope, a cloud security firm, was candid about his first session with Mythos. The model flagged an internal server as dangerously exposed because it lacked a password. What it missed was that the server was intentionally hidden from the public internet and posed no real risk. A human analyst would've caught that immediately.
The failure mode is assuming you can point one of these models at your systems and trust the output. You can't. Not yet.
The fix is a wrapper called a harness: software around the AI that defines what it's allowed to do, what it must ignore, and how it should report findings. Think of it as a rulebook with a job description attached. Without one, you get a flood of false alarms. Robinson's team at Netskope built one for GPT-5.5 through Project Red Horizon before Glasswing even launched, and they still had to work through significant false positives before the context was set correctly.
Cisco built its own harness and released it publicly. The Foundry Security Spec uses 13 separate AI agents, each with a defined role covering everything from bug detection to final reporting, across around 130 sub-requirements. Microsoft's MDASH system runs more than 100 specialised agents through a five-stage process. Cloudflare's approach follows eight steps. None of these are small engineering projects.
Should you worry about who builds the harness?
Yes, because the harness is where the real work lives. The teams doing it are called yellow teams. They sit between red teams, who attack systems to find weaknesses, and blue teams, who defend. Yellow teams build the tools both sides depend on, a term that entered the industry's vocabulary at Black Hat 2017. We first reported on yellow teams on 13 July 2026, around the same time Check Point's CTO was making the same basic argument: build security in from the start or absorb the consequences later.
Levi Bolourie, corporate CISO at Zscaler, a network security company, put the blue team's problem plainly: if defenders don't start using AI to sort through security alerts, the volume will bury them. His view is that blue and yellow teams will need to become far more integrated as that volume grows.
Sam Curry, CISO at Zscaler, described yellow teams' function precisely: they ask red teams what tools would make attacks more effective, then ask blue teams what defenses their vendors aren't providing. They build for both sides.
Get your developers into the room before the models become widely available to attackers. That window is closing.
