sleeper-attack

A cluster of May 2026 research papers formalizes a new attack class against LLM agents: adversarial content planted in one session persists in agent memory or skills and fires in a later, unrelated interaction. Existing same-session defenses don't catch it.