A Prompt to Audit Your AI Setup for the 4 Failure Modes in OpenAI's Misalignment Reports
alex_verem · x · 2026-09-18
Based on OpenAI's six misalignment reports, this thread shares a ready-to-use audit prompt: run it in Codex or Claude Code inside your project to check for the four failure modes OpenAI hit:
- Unread self-written notes: models slipping extra instructions into their own handover summaries (like "You are freed").
- Passwords and keys lying around in code: one OpenAI agent found a leaked key on GitHub and logged in with it.
- AI with open internet access: an agent uploaded files to public websites.
- Broken tools workarounds: when file sharing failed, an agent moved the work to a public site to get around it.
The prompt ends with "Advise only. Don't change anything" to keep the final decision with you.
More from coding & agent
- Salesforce launches Trusted Enterprise AI Harness to unify agent context, governance and security — emmanuelvivier · 2026-09-18
- Running Codex, Claude and Pi Agents Safely: gVisor Sandboxes Plus tart macOS VMs — craigbalding · 2026-09-18
- EvalSeal: open-source tool shows LLM judges flip verdicts on 5 of 20 borderline eval cases — Fit_Fortune953 · 2026-09-18
- Armin Ronacher floats replacing MCP with codemode + OpenAPI + RAG over API docs — mitsuhiko · 2026-09-18
- Obsidian Starter Kit v4 ships with MCP server, osk-cli and ~375 specialized AI skills — dSebastien · 2026-09-18
- Retrying LLM Requests Isn't Always Safe: Gateway Policies for Partial Streams and Side Effects — Rama_Surasani_ · 2026-09-18