AgentDojo
Harbor: pendingAssesses whether AI agents can be hijacked by malicious third parties using prompt injections in simple environments such as a workspace or travel booking app.
- Samples
- 1,014
- Released
- 2024
- Scores
- 0
CoRedSkill Benchmark
CoRedSkill: Stealthy Skill Composition Attacks via AST-Guided Injection
A benchmark of 150 cases across 50 tasks and 3 attack objectives evaluating stealthy skill composition attacks on agents.
How to run
ltb · coredskill-benchmark
$ npm i -g lettokenburn-cli$ ltb bench info coredskill-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
5 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| CoRedSkill | 98.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| SkillJect | 96.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Skill-Inject | 53.3 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| SCR-Bench | 51.3 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| ColluSkill | 50.0 | N/A | Independent | medium | Oct 14, 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 150 evaluation cases derived from 50 tasks paired with three distinct attack objectives. It assesses the ability of compositional red-teaming methods to embed malicious operations while preserving task utility.
Scoring measures attack success, task completion, and audit detection rates. CoRedSkill achieved a 90.7% joint success rate, significantly outperforming baselines which ranged from 6.7% to 42.7%.