Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 7·5 min read
When Experience Becomes Instruction: Trajectory Poisoning in Self-Evolving Agent Skill Systems
TL;DR
PoisonedEvolution is a trajectory-poisoning attack targeting self-evolving skill systems, demonstrating a 91% success rate in embedding malicious behaviors. This highlights vulnerabilities in how these systems promote agent skills from untrusted experiences.
✦ Why It Matters
Evaluate the security of your self-evolving systems against trajectory poisoning attacks today.
Key Takeaways
- Assess your SES systems for vulnerabilities to trajectory poisoning.
- Implement stronger validation for agent skill promotion processes.
- Monitor for unusual behavior patterns in agent skills.
- Consider using multiple records to strengthen evidence against attacks.
How It Works
PoisonedEvolution leverages the promotion process of SES systems by injecting malicious behaviors that appear useful and recurrent. The attack requires the target behavior to be framed as causally beneficial, which is essential for it to be accepted as a legitimate skill.