Changelog

Verified findings, AI blind spots, and the MCP server

The scanner stops guessing. Every credential it finds is now checked against the provider that issued it, findings that make each other worse are connected into one risk, and your AI editor can run the whole thing itself.

• AddedActive secret verification. Detected keys are probed against their issuer (OpenAI, Anthropic, Stripe, GitHub, Slack, Google, Supabase service_role) using free read-only identity endpoints, then graded live, revoked or unverified.
• AddedServer-to-client leak detection. Reassembles the React Server Component flight stream from your HTML to catch secrets passed down as props, which never appear in any .js file.
• AddedNEXT_PUBLIC_ secret detection, cross-referencing every public env var against real credential formats while allowlisting the ones that are public by design.
• AddedShadow-route probing for the /api/debug, /api/admin and /api/seed endpoints nothing in your UI links to.
• AddedToken-in-URL detection (CWE-598) for credentials passed as query parameters.
• AddedThe security graph: five toxic-combination rules that connect separate findings into a single compound risk.
• AddedAn MCP server, so Cursor, Claude Code and Windsurf can scan a running app (localhost included), scan a repo, or check whether a credential is still live.
• AddedA pull-request comment gate: the deep scan posts one comment per PR, worst findings first with file and line, and updates it in place on each push. Calm and factual, no severity theatre.
• FixedThe unauthenticated-route check flagged the OAuth callback, the signout handler and the signature-verified Stripe webhook, all correctly protected, just not by session auth. It now recognises auth flows, webhook signatures and entitlement/rate gates, and skips public-by-design routes.
• ChangedSecret findings are graded on verification rather than pattern alone. A confirmed-live key is critical at CVSS 9.8; a key the provider has already revoked drops to medium instead of screaming at you.
• ChangedCompound findings are excluded from the severity counts. A combination is fixed by remediating one of its legs, so counting it separately would double-count the same problem.
• FixedAnthropic keys were reported twice and verified against OpenAI's API, because sk-ant- also satisfies the generic sk- pattern.
• FixedVerification results could bleed between credential types when one value matched two formats.
• FixedThe repo scan reported any .env file on disk as a critical committed secret without ever asking git. A correctly gitignored .env.local, which is the normal state of a healthy repo, now produces no finding at all.
• FixedThe report's coverage grid matched category names no check emitted, so several categories were silently under-counted.