Our no-hardcoded-credentials rule fired 842 times on vercel/ai. The peer plugin fired 380. I assumed we had better recall — until I sampled. 807 of the 'extra' findings were TypeScript union-type literals, error class names, and the string 'test'. The real number of hardcoded credentials was zero. Here's how a context-blind regex becomes a context-aware detector — and why AI assistants keep regenerating the exact strings that fool it.
A credential scanner that reports 842 secrets in a codebase with zero hardcoded secrets isn't cautious. It's broken. Mine was, and it took a peer plugin reporting less than half my count to make me check.
The flagship rule secure-coding/no-hardcoded-credentials runs alongside eslint-plugin-no-secrets/no-secrets on vercel/ai (the AI SDK) as part of our ILB-Flagship bench. Both rules see the same source. Findings count:
Rule
Findings on vercel-ai
secure-coding/no-hardcoded-credentials (ours)
TL;DR AI editors keep hardcoding API keys and secrets directly into source files That...
TL;DR AI editors hardcode API keys, tokens, and JWT secrets straight into source because...
SearchLeak and a three-CVE LiteLLM chain broke the same AI trust boundary in two weeks. A 5-check audit maps each gap to a CVE, a…
A static AST scan of three open-source TypeScript agent codebases found 669 tool calls with real side effects. 553 had no guard…
Your AI assistant just wrote 200 lines of authentication middleware. It looks clean. It passes the...
Two July 2026 disclosures show AI coding agents executing planted payloads and leaking private repos, all through context the…
