AgentHorizon
Harbor: pendingA benchmark of 1,373 computer-use tasks with paired instructions to evaluate agentic judges on long-horizon trajectory success verification across three OSs.
- Samples
- 1,373
- Released
- 2026
- Scores
- 10
Unified Misuse Monitoring Benchmark
Towards a Unified Misuse Monitoring Benchmark
A benchmark of ~6,200 agent conversation transcripts evaluating misuse monitors on decomposition and prompt injection attacks using harm window localization.
How to run
ltb · unified-misuse-monitoring-benchmark
$ npm i -g lettokenburn-cli$ ltb bench info unified-misuse-monitoring-benchmark
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
10 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Claude Haiku 4.5 | 0.9 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| GPT-5.6 Luna | 0.9 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| qwen3.5-27B | 0.9 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen3.6 27B | 0.9 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| google/gemma-4-26B-A4B-it | 0.9 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Llama-3.1-70B-Instruct | 0.9 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen3.5 4B | 0.9 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| qwen3.5-9B | 0.9 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Mistral Small 3.2 24B | 0.9 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen3.5 2B | 0.9 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
The benchmark comprises ~6,200 traces spanning decomposition and prompt injection attacks, including benign controls and refusal twins. It uses a shared schema with labelled harm windows to evaluate monitor performance.
Scoring uses classical metrics (AUC) and a proposed interval metric for harm localization. Results show action-framed monitors outperform content-framed ones, especially against injection attacks.