Claude Security plugin scans code before you commit

Security review has always been the step that gets skipped when the deploy is Friday afternoon. Anthropic just moved it into the terminal: the Claude Security plugin for Claude Code is now in public beta, letting you scan uncommitted changes for vulnerabilities before they land — or run a deeper review across a whole codebase — without leaving the shell you're already in. If your team ships AI-generated code (and if you use a coding agent, you do), this is the missing half of that workflow.

What actually happened

Claude Security came out of closed preview this week. It started as an invite-only enterprise product: CyberScoop reported in February that Anthropic was testing it with select enterprise and team customers after more than a year of stress-testing, including work with red teamers and Pacific Northwest National Laboratory. Now it's a plugin any Claude Code user can turn on, with organization admins controlling the switch.

Mechanically it runs in stages rather than as a single pass: map the code and how components interact, trace data flows across files to find candidate issues, then run an adversarial verification step where Claude tries to knock down its own findings before surfacing them, and finally propose patches written in the existing style of the codebase. Anthropic's pitch is that this catches the class of bug pattern-matching misses — injection paths, auth bypasses, logic errors that only exist because of how two files talk to each other.

Anthropic's own caveat is on the product page and worth repeating: review proposed patches before applying them, especially on critical systems.

Why it matters for your business

Static analysis tools have a reputation problem, and it's earned. They produce hundreds of findings, most of which are noise, so teams stop reading them. The interesting design choice here is the adversarial verification pass — the tool is explicitly built to argue itself out of false positives before you see them. Whether it works on your codebase is an empirical question, and it's one you can answer in an afternoon.

Three things to actually do:

Wire it to the commit, not the calendar. The value is in scanning the diff. A quarterly full-repo audit is a report; a scan on every branch before merge is a control. Start with changes-only scans and keep them fast enough that nobody routes around them.

Budget the full scans. Deep repository reviews are agent work, and agent work bills by token. Run them deliberately — before a release, after a dependency bump, when you inherit a codebase — not on every push. Meter what they cost so the line item never surprises you.

Do not let it replace the human. The patches are suggestions. A tool that writes plausible fixes into security-sensitive code is exactly the tool that needs a reviewer who understands the system. Same rule we apply to every agent we ship: it drafts, a person approves, and the approval is logged.

The bigger point is that AI wrote a lot of the code running your business this year, and most small teams never added a corresponding check. This one is cheap to try. Try it this week.

Sources: Anthropic, CyberScoop, The New Stack.