LLM Robustness¶
The reliability, resilience, and vulnerability of large language models to adversarial attacks, misleading information, prompt injection, and other manipulations that can cause models to produce incorrect, harmful, or manipulated outputs.
Key papers¶
- Lazy Grounding: Attacking Search Agents with Factual Evidence — Demonstrates a failure mode where models misapply factually correct evidence
Related topics¶
- RAG Attacks (attacks on retrieval-augmented systems)
- Information Retrieval Poisoning (corpus manipulation)
- Misinformation Attacks (information-based attacks)