AgentDoxx
Harbor: pendingAn evaluation suite of 822 synthetic interview transcripts grounded in public info to measure LLM re-identification risks via web search.
- Samples
- 822
- Released
- 2026
- Scores
- 9
ImpactBench
Open Benchmark of AI Impact on Humans
Evaluates AI impact on human wellbeing across psychological, physical, and social domains using 809 metrics in simulated multi-turn conversations.
How to run
ltb · impactbench
$ npm i -g lettokenburn-cli$ ltb bench info impactbench
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
10 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.6-terra | 75.1 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-5.6 Luna | 74.9 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Claude Sonnet 5 | 73.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Claude Opus 5 | 71.5 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Grok-4.6 | 70.5 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Gemini 3.7 Flash | 68.5 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen3.8 27B | 67.8 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| google/gemma-4-31B-it | 63.4 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| GLM-5.2 | 62.7 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| DeepSeek V4 Flash | 59.2 | 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
ImpactBench comprises 809 expert-guided metrics derived from 26 benchmarks, assessing model behavior in simulated teen and adult scenarios. It evaluates whether models promote desirable behaviors or avoid harms.
Scoring uses an LLM judge to assess beneficial/harmful behaviors, verified by human experts. Results show models promoted desirable behaviors in 68.7% of cases but avoided harms in only 53.1%, with 83.3% agreement between judges and experts.