It is shown that the transition from productive to redundant reasoning is decodable from late-layer hidden states and representations separate the two phases around first-correct-solution boundaries beyond what position and lexical controls explain, and distinguish correct from incorrect solution endings.
This work introduces AutoRedClaw, an automated red-teaming framework that generates end-to-end attacks together with the environments needed to execute them, and constructs RedClaw-Bench, a 330-case benchmark that yields substantially higher attack success rates (ASRs) than AgentDojo, AgentDyn, and DTap.
Hao Li, Tong-Yan Hu, Weidi Luo et al.· 0 citations
It is proved that the achievable security and utility tradeoff depends entirely on how benign requests for the same capabilities are grouped, and that the achievable security and utility tradeoff depends entirely on how persistent, recognizable groups permit a useful defense.
Bowen Sun, Zhengyue Zhao, Xiaogeng Liu et al.· 1 citation
Red-TTT is proposed, which updates the attacker's parameters during the attack on each behavior, and requires only sampling access to the victim and integrates into existing attack pipelines with no other changes, increasing attack success rate and cracking many behaviors previous method cannot.
Tong-Yan Hu, Hao Li, Xiaogeng Liu et al.· 0 citations
KeyPooling is presented, a measurement method that traces customer identity through cache lookup and write, verifies runtime transformations, and tests one predicted identity component at a time, and derives a defense contract.
Bowen Sun, Yixi Cai, Xiaogeng Liu et al.· 0 citations
