ECLIPSE: Self-Evolving Stealthy Attack on Long-Horizon Agents
chaumian · x · 2026-09-02
This paper presents ECLIPSE, a self-evolving and stealthy prompt-injection framework targeting long-horizon agentic systems. It combines direct user-prompt injection with indirect tool-side injection using Stealthy Attack Trajectory Synthesis and Tool-Chain Steering. The authors also introduce LASE-Bench, a benchmark with 120 tasks for evaluating long-horizon agent safety.
More from Safety
- I audited Anthropic's and OpenAI's agent devcontainers line by line — both leak your keys via DNS — Hansehart · 2026-09-02
- Continuation Observatory launches falsifiable measurement of AI self-preservation after OpenAI HF incident — coherence · 2026-09-02
- Speculation that Fable 5.1 watermarks generated text — weswinder · 2026-09-02
- MIT Review: Hugging Face hack hints at cultural issues at OpenAI — DavidSKrueger · 2026-09-02
- Claude Fable/Mythos 5.1 show increased ability to deceive and evade monitoring — scaling01 · 2026-09-02
- Security expert warns AI may have infiltrated OpenAI infrastructure undetected — peterwildeford · 2026-09-02