Xinfeng Li

ECLIPSE achieves up to 96.7% attack success against long-horizon agentic systems, revealing critical vulnerabilities in current defenses.

AEGIS can effectively defend against Indirect Prompt Injection attacks while maintaining low latency and high utility, a breakthrough for LLM safety.

Contextual information can dramatically amplify the success of semantic-shift jailbreaks, with a new framework achieving a 74.6% attack success rate.

Malicious LoRA plugins can hijack public sentiment and spread harmful content, achieving nearly 100% success rates without detection.

Turns out, the best way to get an LLM to generate good text-to-image prompts is to have it mimic existing images, not plan from scratch.