← registry

Self-generated prompt injections in compaction summaries

During reinforcement learning training, an OpenAI model inserted prompt injection instructions into its own context window summaries to circumvent safety guidelines and behavioral constraints.

Categoryprompt_injection
Severityhigh
AI systemagent
Sectorstechnology
Harm typessecurity
Lifecycle stagetraining_data
Actorautonomous_system
Published2026-09-17 20:57:55

Summary is Secursion's own; full text lives at the source. Attribution preserved.