What Claude Security gets right — and what it misses

Anthropic just opened Claude Security to public beta. It scans your codebase, finds vulnerabilities, and writes the patch — all in one sitting, no ticket queue. That's a genuine capability worth taking seriously.

But if you're a security leader reading the launch coverage and thinking "okay, this changes things for us," look a little deeper.

The part nobody's arguing with

Let's be honest about what Claude Security actually does well, because the instinct to dismiss new AI security tools is its own kind of lazy.

Traditional static analysis tools pattern-match against known vulnerability signatures. Claude Security reads code the way a security researcher does — tracing data flows, examining how components interact, looking for things rule-based tools are structurally incapable of seeing. Each finding goes through a validation pipeline before it reaches anyone: confidence ratings, severity, reproduction steps, a suggested fix.

That's better than most automated scanning tools, which are historically so noisy that security teams learn to deprioritize everything they flag. Anything that improves signal-to-noise is genuinely useful. For stretched teams with more to do than capacity to do it, that compression matters.

"Finding a vulnerability is a technical problem. Fixing it is a process problem."

Why the governance layer still matters — more than ever

An AI-generated patch still needs someone to review it. Someone who understands whether the fix is actually correct — whether it closes one hole while quietly opening another. Someone who can make the call on deploying a model-generated change to production with the right change management controls, against the right risk tolerance, in that environment, on that system, at that moment.

That governance layer does not compress. It doesn't disappear because a model wrote the suggestion. If anything, it matters more when AI is generating the fix — because the model doesn't know your environment, your architecture, your downstream dependencies, or your risk appetite.

The gap between suggestion and safe deployment is where most organizations are actually exposed. And it's a process gap. No tool closes it for you.

Your attack surface is not your codebase

This is the part of the Claude Security conversation that isn't happening, and it should be.

Claude Security scans code. That's the scope. That's it.

When we're doing penetration testing, we routinely find the real exposure sitting somewhere entirely different — a misconfigured cloud environment, weak or reused credentials, a poorly segmented network, a forgotten dev environment someone spun up six months ago with production access still attached. The application code is clean. Everything around it is wide open.

The majority of breaches still trace back to exactly these kinds of issues. Unpatched systems. Credential problems. Poor network segmentation. Not sophisticated application-layer exploits. AI-powered scanning gets faster at finding code vulnerabilities — it doesn't touch the rest of the attack surface where most of the actual risk lives.

An organization can have reasonably clean application code and still be in serious trouble. A tool that scans one layer doesn't tell you anything about the others — and treating it as though it does is the kind of shortcut that looks reasonable in a board presentation and falls apart in an incident.

So what does this actually mean?

Claude Security is useful. For teams with mature code review processes, strong change management, and the human capacity to properly validate and act on findings, it's a meaningful addition to the toolkit.

But "does this tool work" and "is our security program structured to handle AI-driven threats" are genuinely different questions. A lot of organizations are currently focused on the first and not asking the second.

The second question requires looking at the whole program — application layer, cloud posture, network architecture, credential hygiene, third-party risk, and the governance processes that sit across all of it. It requires being honest about whether your program is structured to operate at the speed AI-driven threats actually move, not just whether you've deployed a scanner.

AI is raising the bar for what prepared looks like. Anthropic's own launch messaging says it directly: the window between vulnerability discovery and exploitation is shrinking. Defenders need to move faster. Moving faster on one layer of the stack, while leaving the rest unexamined, isn't the answer.

See where your program actually stands

The OSec Mythos Readiness Assessment looks at the whole picture — application layer, cloud posture, network, third-party risk, and governance — based on the CSA framework. Takes three minutes.

Take the Assessment