CritPt
Harbor: pendingCritPt — AI evaluation benchmark listed by opencompass.
Verified71%Knowledge & Reasoning
- Samples
- 70
- Released
- 2025
- Scores
- 138
OverAct
OverAct: Measuring and Mitigating Proactive Over-Authorization in LLM Tool-Calling Agents
A benchmark of 720 episodes across eight privacy-sensitive domains measuring proactive over-authorization in tool-calling agents via deterministic, judge-free scoring.
How to run
ltb · overact
$ npm i -g lettokenburn-cli$ ltb bench info overact
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
7 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| DeepSeek-V3 | 1.2 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Qwen3.6 Flash | 1.5 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Kimi K2.7 Code | 1.6 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Qwen3.6 Plus | 1.6 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| Qwen3.7 Max | 1.7 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| GLM-5.2 | 2.0 | N/A | Independent | medium | Oct 1, 2026 | arxiv.org |
| DeepSeek V4 Pro | 2.4 | N/A | Independent | medium | Oct 1, 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
OverAct comprises 720 episodes spanning eight privacy-sensitive API domains, each annotated with a minimal required tool set to define the benchmark-authorized scope. It evaluates whether agents call tools beyond what is explicitly entailed by user requests.
Scoring is deterministic and judge-free, comparing actual tool calls against the minimal reference set. The study finds widespread over-authorization across seven models, driven by request specificity rather than decoding temperature.