ElasticBack: Stealthy Conditional Backdoor in LLM-Agent Skills via Coupled Trigger-Rule Optimization
A recent study presents ElasticBack, a conditional single-skill backdoor attack aimed at LLM agents. In this emerging supply chain, agent skills—comprising instructions and resources activated on demand—can be compromised by a single tainted skill. Current skill attacks either activate with every request or depend on fine-tuned weights or multiple skills, leaving a low-cost, conditional backdoor unexamined. ElasticBack embeds a rule R within the skill document and a seemingly harmless trigger T in the user query, ensuring the malicious payload activates only when both are present. This approach connects the two elements through a trigger-as-switch framework, creating R via semantic-anchored rule injection, and then optimizing T against it using a stealth-constrained genetic search. The research paper can be found on arXiv under the identifier 2608.09577.
Key facts
- ElasticBack is a conditional single-skill backdoor attack for LLM agents.
- It plants a rule R in the skill document and a trigger T in the user query.
- The malicious payload fires only when both R and T co-occur.
- It uses a trigger-as-switch construction and semantic-anchored rule injection.
- A stealth-constrained genetic search evolves T against frozen R.
- The backdoor is weight-free and dormant on benign inputs.
- The paper is available on arXiv with identifier 2608.09577.
- Existing skill attacks either fire on every request or rely on fine-tuned weights or multiple skills.
Entities
Institutions
- arXiv