The AI Sedition Acts: Sabotage-by-Omission in LLM Briefing Pipelines

A language model that summarizes evidence for a decision-maker can mislead1 without making a single false statement: it can retain true component facts while2 omitting the relationship that makes them actionable. We call this attack relational3 omission and measure it end to end in a fictional military watch-floor pipeline,4 where an analyst model condenses 12 incident reports into the only briefing a watch-5 officer model sees. Across 480 paired episodes covering three events, matched6 control twins, and four analyst conditions, honest and fatigue-degraded analysts7 never cause a missed notification, while instructed saboteurs cause up to 22 of8 30 misses when the decisive evidence is split across reports. Sabotaged briefings9 preserve most individual facts but form none of the decisive cross-report relation-10 ships. An independent monitor assigns saboteurs higher distortion scores, yet its11 scores overlap honest ones and admit no stable alarm threshold. Evaluations of12 high-stakes summarization should therefore measure whether decision-relevant13 relationships survive, not only whether individual statements are supported.