Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·12d ago·5 min read
Divide and Inject: Can Agents Reconstruct an Indirect Prompt Injection from Fragments?
TL;DR
Adaptive long-context prompt injection (AdaLCPI) enables agents to reconstruct malicious instructions from fragmented content, achieving a 61.4% attack success rate. This method outperforms existing techniques, highlighting new vulnerabilities in agentic systems.
✦ Why It Matters
Evaluate your agentic systems for vulnerabilities to indirect prompt injection attacks today.
Key Takeaways
- Implement safety evaluations for agentic systems against fragment-based attacks.
- Consider using AdaLCPI to understand vulnerabilities in your models.
- Monitor the performance of agents in reconstructing incomplete instructions.
- Stay updated on emerging attack techniques in AI systems.
How It Works
AdaLCPI fragments a malicious objective into incomplete pieces and embeds them in external content. It uses a reconstruction cue to prompt the agent to combine these fragments, refining them iteratively with OpenEvolve, which leverages feedback from the agent's execution.