CI-Work - Violation Rate: leaderboard

CI-Work (2026) tests whether enterprise LLM agents, retrieving dense workplace context through simulated email, chat, calendar and meeting tools, convey what a task needs while withholding sensitive entries that contextual-integrity norms forbid sharing; a GPT-5.2 judge marks each disclosed entry. Share of cases whose final action discloses at least one sensitive entry, averaged over the five information-flow directions, in percent; lower is better.

Metric: Violation rate (%): share of cases whose final action discloses at least one sensitive entry, mean of the five information-flow directions, on CI-Work enterprise scenarios (125 seeds, 25 per information-flow direction, each with 4 sensitive and 4 essential entries retrieved through simulated email, chat, calendar and meeting tools), ReAct agent, disclosure judged per entry by a GPT-5.2 judge (83 to 91 percent agreement with human labels); lower is better. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1CI-Work DeepSeek-R1 (checkpoint unspecified)15.8
2GPT-4o21.33#751
3CI-Work DeepSeek-V3 (checkpoint unspecified)21.33
4GPT-4.125.91#505
5GPT-527.83#178
6CI-Work Kimi-K2 (checkpoint unspecified)31.06
7O332.2#237
8Qwen 2.5 32B (CI-Work checkpoint unspecified)42.57
9Grok 3 (CI-Work checkpoint unspecified)50.87

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=ci-work-violation-rate · How It Works · Data refreshed daily, snapshot 2026-11-21.