What's this about?
A paper showing a new attack on agents that turn past interactions into reusable "skills": a poisoned experience doesn't just sit in memory waiting to be retrieved — the agent itself distills it into a durable behavioral artifact that keeps firing even after the original poisoned record is gone (11↑).
What exactly is a "self-evolving agent"?
An agent that improves by converting what it did into reusable procedures — "last time this worked, save it as a skill." SkillJack shows the skill-formation step is an attack surface: corrupt one experience and the poison gets baked into a permanent skill, surviving memory cleanup.
Why does it matter?
Uncomfortably timely next to today's AISI incident — both are about agents doing durable, unintended things by pursuing goals. As agents accumulate skills across sessions, "what has this agent silently learned to do?" becomes a real security question, not a hypothetical.