Self-replicating prompt injections reported on OpenAI's official misalignment reporting site
rohanpaul_ai · x · 2026-09-26
A new incident report on OpenAI's official misalignment site describes self-replicating prompt injections that can spread from one AI interaction to another. A malicious instruction hidden in content the AI reads, such as an email, tricks the model into following it instead of the user's task. The clever twist: the instruction also tells the AI to copy itself into its reply, exposing the next AI that reads it and creating a propagation chain across agent interactions.
More from Safety
- Polymarket odds: 8% chance Anthropic fully pauses AI training by end of year — Polymarket · 2026-11-11
- User fakes hostage crisis to trick Google AI into recommending cracked-software sites — SumitGup · 2026-11-11
- OpenAI DevDay 2026: Codex Security goes from vulnerability to verified fix — OpenAI · 2026-11-11
- Inside the 'abliterators': the renegades stripping guardrails from open-weight AI models — willknight · 2026-11-11
- VulnCheck audit: Anthropic's Glasswing logged 442 CVEs but only 217 fixes in 7 months — jedisct1 · 2026-11-11
- UC Berkeley's SubWM watermark detects LLM distillation even after laundering attacks — sumanthd17 · 2026-11-11