New Technique Medium TD-2026-0119 For AI Eng, Red Team, CISO AI-assisted
SCoRe context compaction lifts agent exploit capability past stop responses on ExploitBench
Language model agents on ExploitBench often stop with budget remaining, and additional verified gains appear after the first stop response, though they diminish over longer runs. The authors propose SCoRe, which compacts context at stop bursts and marks impossibility judgments as unproven. With SCoRe, MiMo reaches a 67.0% capability coverage score at 2,000 turns, versus 32.4% under the original stopping policy.
Why it matters
Organizations evaluating autonomous agents for vulnerability exploitation may underestimate capability if they treat early stop responses as final limits. Context management choices can change measured offensive capability substantially.
Do this
- Treat agent stop responses as provisional in internal evaluations and continue runs with budget remaining before recording a capability ceiling.
- Audit context summaries used by agent scaffolds for retained claims that a capability is impossible, and flag such claims for re-verification.
- Track the proportion of stop responses within a single context as a monitoring signal for agent behavior in long-horizon tasks.
- Re-benchmark exploitation-capable models after any change to context compaction or summarization logic, since results can shift materially.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“We find additional verified gains after the first stop response, but gains diminish over longer execution.”
“With SCoRe, MiMo reaches a capability coverage score of 67.0% at 2,000 turns, exceeding both 32.4% under the original stopping policy”
Sources
Affected / related
SCoRe context compaction, Sequential test-time scaling, MiMo-V2.6-Pro, DeepSeek-V4-Flash-0731, GPT-5.5
Published 2026-11-17 14:58 UTC · updated 2026-11-18 04:35 UTC · AI-assisted summary, checked against quoted sources · How we score