Answers are grounded in this paper and cite their sources.
Large language models (LLMs) can generate cyber-defense plans whose action order is unsafe for an exposed critical server. We evaluate one implemented repair-layer instance: a predicate that moves patch(v) first only when an uninfected, non-isolated, unpatched critical server has at least two infected, unisolated neighbors under critical_targeted mode. In 720 matched enterprise episodes, repair raised the external score by 11.46 points without retrieval and 23.13 with retrieval; both gains arose entirely from 60 critical-star tuples, with zero gains in the other 120. Both contrasts remained positive across 84 alternative score-weight vectors. Under a common 12-transition horizon, a rule-only shield implementing the same predicate scored 91.09, above the repaired retrieval plan at 68.11; the complete system also trailed the fixed rule in the original aggregate experiment (58.25 versus 67.85). Runtime shielding and plan repair produced identical external outcomes in 180 matched episodes, as expected because they enforce the same predicate. Dynamic LLM reward shaping scored 54.50 versus 53.64 for Q-learning while requiring 3143 API calls. CAGE used a distinct process-alert adapter. The evidence supports a scenario-specific repair-layer instance, not a general advantage of LLM planning over deterministic rules.
No takes yet. Share an insight, caveat, or question.
Xiang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: