CRIME
Compositional Risk Induction via Multi skill Execution
A benchmark of 4,000 public skills across eight cybersecurity behaviors to evaluate composition-induced vulnerabilities in LLM agents.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info crime
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
2 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-4o-mini | 14.4 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| DeepSeek V4 Flash | 37.9 | 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 runsAbout
The benchmark comprises 4,000 public skills spanning eight cybersecurity behaviors, designed to test if benign skills compose into malicious actions. It evaluates both direct composition and adaptive refinement attacks within a sandboxed environment.
Scoring relies on examining environmental consequences in the Skill Reaction Chamber to determine if target malicious behaviors occur. Experiments show that individual skill inspection fails to capture these multi-skill security risks.