Researchers extend AI agent vulnerability exploitation past its stopping point with SCoRe
An arXiv paper reports that language model agents often stop exploiting software flaws with budget to spare, and that a context-management method called SCoRe recovers much of that untapped capability.
In plain words
AI agents that try to exploit known software flaws sometimes quit early, even when they still have computing budget left. The paper's authors found that letting the agents keep going, while tidying up their memory of what has and has not been proven, raised their measured success. This suggests a stop is not necessarily the agent's real limit.
What was announced
Writing on arXiv, the authors study vulnerability exploitation, where language model agents must build working exploits from known software flaws. Using ExploitBench with DeepSeek-V4-Flash-0731 and MiMo-V2.6-Pro, they report that agents frequently stop while budget remains, and that continuing past the first stop yields additional verified gains, though these shrink as execution lengthens. They link less productive continuation to two patterns: a growing share of stop responses within one context, and earlier stopping tied to summaries that carry impossibility judgments about capabilities the agent has not achieved.
To address this, the authors propose SCoRe, which compacts the context when stop responses cluster and marks those impossibility judgments as unproven. They report that MiMo-V2.6-Pro with SCoRe reached a capability coverage score of 67.0% at 2,000 turns, above 32.4% under the original stopping policy and above the 58.6% they computed for GPT-5.5 from ExploitBench's 300-turn reference results on the same tasks. In their view, stopping need not mark an agent's exploitation limit, and continued execution with context management can elicit substantial further capability.
What it does not show
- The evidence is the paper's abstract only; the full article's methods, baselines and caveats were not available for review.
- The reported scores come from the authors' own evaluation and have not been independently confirmed.
- The GPT-5.5 figure was computed by the authors from ExploitBench reference results rather than measured under identical conditions.
For security and IT teams
- Scope
- The findings apply to long-horizon vulnerability exploitation tasks on ExploitBench with the specific models named, not to agent behaviour generally.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.