Skip to content

Stopping Is Not the Limit: Understanding and Eliciting Agent Capabilities in Long-Horizon Vulnerability Exploitation

Stopping Is Not the Limit: Understanding and Eliciting Agent Capability in Long-Horizon Vulnerability Exploitation

arXiv: 其他 AI 安全论文· arXiv:2611.22888· Deokhyung Kang, Minjae Gwon, Sangdon Park·Indexed by this site · 2 days ago· Originally publishedEvaluation· Its place in the scan record →
AI summary

In the ExploitBench exploitation test, the researchers found that language model agents still had verifiable capability gains from continuing execution after stopping, and propose a context management method, SCoRe, to elicit them. The experiments used DeepSeek-V4-Flash-0731 and MiMo-V2.6-Pro; agents often stopped before exhausting the budget, and gains diminished as execution lengthened. SCoRe compresses context when stop responses are dense and marks impossible judgments about unachieved capabilities in summaries as unverified, outperforming summary-based continuation and other evaluated context management methods. With SCoRe, MiMo reached a capability coverage score of 67.0% at 2000 rounds, higher than its 32.4% under the original stopping policy and than GPT-5.5's 58.6% computed from the ExploitBench 300-round reference results on the same batch of tasks.

Read originalarxiv.org