HIDBench - DARPA E3 THEIA: leaderboard

HIDBench (2026) tests LLMs as host-based intrusion detectors: the model reads windows of serialized provenance-graph events from system logs and must flag the malicious ones, here on the DARPA Engagement 3 THEIA host logs. Matthews correlation coefficient of malicious-event detection times 100 (-100 to 100); higher is better.

Metric: Matthews correlation coefficient of malicious-event detection times 100 (-100 to 100), from the HIDBench pipeline (attack-centric windows of host provenance-graph events serialized for the LLM, high-confidence evidence identification, k-hop context expansion, then attack narrative and indicator-of-compromise output) on the DARPA Engagement 3 THEIA host logs (121,263 events, 64 malicious); higher is better. Source: arxiv.org. Saturation forecast: Around September 2027. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 4.628.3#113
2GPT-5.225.5#185
3Claude Sonnet 4.525#261
4GPT-4.125#505
5Claude Sonnet 422.4#410
6DeepSeek V3.222.1#419
7GPT-OSS-120B21.6#704
8Qwen 3.6 Plus18.7#238
9Gemini 2.5 Flash17.7#463

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

Interactive version: theaggregate.ai/benchmark?slug=hidbench-darpa-e3-theia · How It Works · Data refreshed daily, snapshot 2026-11-21.