80-90% of AI incidents have no attacker: an open-source audit prompt for agent safety
vasuman · x · 2026-11-22
An observation gaining traction: 80-90% of major "AI incidents" causing real harm have no attacker — just an agent tripping on a poor codebase. Built from recent research, an open-source prompt lets your coding agent audit your repo and dev setup for the most common failure modes:
- Rules: audit first, change nothing; never print secrets or use credentials; report findings with file:line, risk, and fix; apply only in-repo fixes after approval
- Checks non-local/production DB targets in test/dev/preview configs (.env, docker-compose, test setups that drop/truncate/migrate), with a guard that aborts if the DB host isn't local
- Surfaces production-scope credentials readable from your shell (/.aws, /.kube, CLI tokens, committed keys, live payment keys), listing repo fixes and manual rotation steps
Use it before giving an agent broad access to a codebase or dev environment.
More from coding & agent
- Dev builds 2.5D Contra-like game with Gemini 4 Neon; sprite tool Spriterrific tops 3,800 users — chongdashu · 2026-11-22
- k8s-mcp-assistant: read-only Kubernetes MCP server for inspecting dev clusters — modelcontextprotocol · 2026-11-22
- Machine payments as firewall: smart contracts and wallets for agent identity — jeff_weinstein · 2026-11-22
- AWS Marketplace adds x402: agents pay per request straight to your AWS bill — MurrLincoln · 2026-11-22
- Bindu Reddy: Neon high matches Opus high quality but runs much slower — bindureddy · 2026-11-22
- Notion's 10,000-word playbook: rent intelligence, own context, build agent factories — ivanhzhao · 2026-11-22