Articles
Engineering writing on static analysis, ESLint, security, and AI-native developer tooling.
Filtered by tag #ai · clear filter
Dec 20, 2025I Linted 12 AI-Shaped Agent Tools. All 12 Were Missing an Input Schema.I ran eslint-plugin-vercel-ai-security over 12 tool definitions written the way AI assistants wire them: 12/12 had no inputSchema and 6 destructive tools (deleteUser, transferFunds, executeCommand…) had no confirmation gate. Five CWE-tagged ESLint rules bound an AI agent's agency at write-time — and one ungated tool still slipped past.#ai#security#javascript+1008 min0
Dec 19, 2025I Linted Every OWASP LLM Category Against a Real Vercel AI SDK App. 8 Fire. 2 Are Vendor Theater.A real mapping of the OWASP LLM Top 10 (2025) to eslint-plugin-vercel-ai-security: which categories a CWE-tagged rule genuinely catches in source — and why each one survives normal code review.#ai#security#javascript+19 min
Dec 19, 2025Every AI Model Ships a Vulnerability in Half the Code It Writes. Prompt Injection Has 3 Faces in Your Vercel AI SDK App — Here's the ESLint Rule for Each.Across 700 AI-generated functions from five models (measured 2026-02-09), every model shipped a vulnerability in 49–73% of them. Prompt injection has three faces in a Vercel AI SDK app: unvalidated input, a leaky system prompt, and unconfirmed tool calls. Three CWE-tagged ESLint rules catch each at write-time — plus the hardened pattern they want.#ai#security#javascript+18 min
I Let AI Agents Rebrand My Entire OSS Ecosystem in One Night. Two Pipelines Were Silently Dead.Merged is a feeling. Live is a measurement. 13 PRs across 5 repos in one night of parallel AI agents — and two delivery pipelines that had been silently dead for weeks.#ai#webdev#security+14 min
innerHTML Has Five Doors. Most Reviews Only Watch One.A DOM sink is easy to grep for. The hard part is that one sink has many sources, and the source is where the trust decision actually lives — not the assignment.#ai#webdev#security+16 min