Xing Zheng

The account motivates complementary defenses: strengthen resistance and recovery alongside prevention of spontaneous deviations, and audit and restrict unintended communication paths that can turn local failures into collective loss of control.

These results show how a latent loss of control can become an external violation: the task objective remains intact, but an executable opportunity can turn a missing control boundary into consequential action.

Achieving high-quality model performance with just 10% of the required labels could revolutionize the scalability of RLVR in large language models.

Extreme fidelity loss reveals critical vulnerabilities in long-horizon tasks that standard accuracy metrics overlook.

Indirect prompt injection can compromise AI systems like DeepSeek Harness, with attack success rates reaching up to 25.5% under certain conditions.

SkillJack reveals that self-evolving agents can unknowingly incorporate malicious skills, making traditional safety measures ineffective against persistent threats.

Despite high rationale identification, search-augmented models struggle with refusal, achieving only 42.9% correct halting on unanswerable multi-hop questions.

EviSD achieves state-of-the-art performance in question-answering tasks by leveraging privileged evidence, outperforming existing methods while maintaining efficiency in response generation.

AI-Infra-Guard reveals that a unified security framework can effectively address the diverse attack surfaces of AI agents, making it a game-changer for AI safety.