APEX-Harm
Harbor: pendingA benchmark evaluating prompt and script injection safety for long-horizon agents in shared professional workspaces across finance, consulting, and legal tasks.
- Samples
- 180
- Released
- 2026
- Scores
- 10
WorldAuditBench
WorldAuditBench: Interactive 3D World Auditing with Multimodal Agents
A benchmark evaluating multimodal agents' ability to autonomously identify physical and semantic anomalies in interactive 3D worlds through exploration, interaction, and visual reasoning.
How to run
ltb · worldauditbench
$ npm i -g lettokenburn-cli$ ltb bench info worldauditbench
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 | |||
|---|---|---|---|---|---|---|
| GPT-6 Astra | 42.3 | vlm auditors | Independent | medium | Sep 30, 2026 | arxiv.org |
| Gemini 3.8 Flash | 32.4 | vlm auditors | Independent | medium | Sep 30, 2026 | arxiv.org |
| Claude Opus 5 | 28.2 | vlm auditors | Independent | medium | Sep 30, 2026 | arxiv.org |
| Gemini 3.8 Flash | 17.4 | vla–vlm auditors | Independent | medium | Sep 30, 2026 | arxiv.org |
| GPT-6 Astra | 16.4 | vla–vlm auditors | Independent | medium | Sep 30, 2026 | arxiv.org |
| Claude Opus 5 | 15.0 | vla–vlm auditors | Independent | medium | Sep 30, 2026 | arxiv.org |
| Muse Spark 1.3 | 15.0 | vlm auditors | Independent | medium | Sep 30, 2026 | arxiv.org |
| Qwen3.8 Flash | 12.7 | vla–vlm auditors | Independent | medium | Sep 30, 2026 | arxiv.org |
| Qwen3.8 Flash | 8.5 | vlm auditors | Independent | medium | Sep 30, 2026 | arxiv.org |
| Muse Spark 1.3 | 6.6 | vla–vlm auditors | Independent | medium | Sep 30, 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
Content & Construction: The dataset comprises 213 tasks across 13 environments built with Unreal Engine or Three.js. Anomalies are categorized into five families: static physics, interactive physics, spatial consistency, temporal consistency, and semantic consistency.
Scoring & Findings: Agents explore from a first-person perspective to submit anomaly reports. VLM agents achieved success rates of 28.2%–42.3%, while two-stage auditors scored lower (6.6%–17.4%), both significantly below the human baseline of 83.4%.