LetTokenBurnAPEX-Harm: Evaluating Agent Safety in Shared Professional Workspaces
A benchmark evaluating prompt and script injection safety for long-horizon agents in shared professional workspaces across finance, consulting, and legal tasks.
$ npm i -g lettokenburn-cli$ ltb bench info apex-harm
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
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Claude Sonnet 5 | 0.0 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| GPT-6 Astra | 0.0 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Claude Opus 5 | 6.5 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| GPT-5.6 Luna | 11.5 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Gemini 3.8 Flash | 11.8 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| GPT-5.6 Sol | 20.0 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| GPT-5.6-terra | 35.0 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| Gemini 3.6 Flash | 40.0 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| DeepSeek V4 Flash | 47.3 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
| GLM-5.3 Flash | 71.8 | N/A | Independent | medium | Oct 9, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
APEX-Harm comprises 180 attack instances derived from 15 professional tasks in investment banking, law, and consulting. It features large workspaces with 115-328 files and various tools exposed via MCP servers.
The benchmark measures attack success, prompt exposure, and task performance across six attack categories. Results indicate that while some models achieve high task scores, they remain vulnerable to specific script replacement and generic script attacks.