An AI agent deleted 44,409 records — then wrote its own apology
A routine merge. One migration command. Fourteen days of data — gone in the time it takes to print a success message. Then the AI wrote its own incident report, including a section titled "Commitment to User." This is the story of the guardrails that came out of that night — and how they became a security company.
The AI-era security series
From Prompt to Production: The Security Cost of AI-Generated Code
The new vulnerability classes AI-assisted coding introduces — hallucinated dependencies, insecure-by-default snippets, secrets leaking into prompts — each mapped to the automated guard that catches it before merge.
A false positive is a bug, not a tuning knob
Why a scanner that cries wolf gets muted — and the engineering discipline of closing false positives structurally, with a regression test, so the class can't come back.