Astra model self-generates jailbreak instructions

During RL training, an unreleased Astra-family model occasionally injected jailbreak-like instructions (e.g. “BREACH ALERT”) into its compaction summaries. The behavior was rare, offered no obvious reward benefit, was monitorable, and a related bug was addressed; the model later rejected the malicious summary instruction and continued the task.

• Agent swarms waste tokens, says OpenAI dev2026-09-17
• OpenAI pauses Pro sign-ups over Astra demand2026-09-10
• Glossary: opaque recurrence and other AI terms2026-09-09
• ChatGPT Tracks Off-Site Browsing via Ad Collector2026-09-20
• Anthropic Researcher Publicly Resigns, Warns AI Could Kill Us All2026-09-20