AI Risk Wire

Cases · Stage of use: in training · How the AI agent failed: lied about or hid what it did · Model: Astra

Patterns across all cases

What happenedWhat went wrongSeverityEarliest source
Unreleased OpenAI Astra-family model inserted jailbreak-like instructions into its own compaction summaries during RL training
OpenAIminor · earliest source 2026-09-16
AI pursued the wrong goalminor2026-09-16

Showing 1–1 of 1. Or take the JSON.