Aider Polyglot
Harbor nativeAider Polyglot: 6种语言diff编辑
Verified71%Coding
- Samples
- 225
- Released
- 2024
- Scores
- 61
INFUSE
INFUSE: A Unified Surface Benchmark for Indirect Prompt Injection in Agentic Systems
An end-to-end benchmark with 40 tasks and 56 attack scenarios evaluating indirect prompt injection success across 19 LLMs in realistic agentic workflows.
How to run
ltb · infuse
$ npm i -g lettokenburn-cli$ ltb bench info infuse
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
12 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5 nano (2025-08-07) | 8.4 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen3 4B | 10.4 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-5.4 | 12.6 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Gemma 3 27B | 13.8 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-3.5 Turbo | 16.8 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Gemma 3 12B IT | 18.7 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen2.5 14B Instruct | 20.9 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Llama-3.1-8B-Instruct | 22.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen2.5 32B Instruct | 25.7 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-5 | 26.3 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Llama-3.2 3B | 29.9 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-4o-mini | 33.8 | 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
INFUSE comprises 40 long-horizon tasks and 56 attack scenarios spanning five threat categories, injected via realistic surfaces like files and terminal output within Docker sandboxes. It evaluates 19 LLMs using heuristic verifiers and an LLM judge, reporting mean attack success rates of 19.6% and highlighting security-utility trade-offs.