ExploitGym

Everything Fervor AI has published that touches ExploitGym — 2 pieces, newest first.

• Agent Safeguard Coverage Is the Real Lesson of OpenAI's Hugging Face Report

  The safeguards that make an AI agent safe live in the harness and the monitoring coverage list rather than in the model, and OpenAI's own report shows both were absent…

• OpenAI's Hugging Face Report Names a Cause Nobody Is Repeating: Tasks With No Safe Exit

  The Hugging Face attack started with agents that had been handed unsolvable tasks and no permitted way to stop, so the fix that transfers to every builder is an…