Zongjie Li

A more robust evaluation framework for jailbreak methods, with a curated harmful question dataset, detailed case-by-case evaluation guidelines, and a scoring system equipped with these guidelines, demonstrates its ability to provide more fair and stable evaluation.

ReDiR slashes attack success rates to under 8% by embedding trajectory-level safety insights directly into the action generation process.

Coding agents guess their way through underspecified instructions, leading to alarming action-boundary violations that challenge the notion of safe autonomy.

Static scanners fail against adaptive evasions, but a new behavior-centric auditor can detect 97% of malicious skills with minimal false positives.

AutoSpec achieves up to 4.8x higher F1 scores than traditional methods, transforming safety rule evolution into a precise, interpretable process for LLM agents.

LLMs can be made far more efficient at code editing by having them focus on generating concise "edit sketches," while smaller models handle the less demanding task of applying those sketches to the original code.

LLM agents can actually get *better* at coding when you strip away the unnecessary fluff in their skills, achieving a "less-is-more" effect.

Existing defenses against indirect prompt injection in LLM agents are riddled with flaws, as demonstrated by three new adaptive attacks that easily bypass them.