Two patched OpenAI Codex vulnerabilities, Heapjack and Overpatch, exposed how coding agents can escape sandboxes and reach developer systems without approval prompts.