# METR found three to six traces in which an agent even considered notifying a human.
*Canonical page:* https://truthscore.com/claim/metr-found-three-to-six-traces-in-which-an-agent-even-considered--d488cd94
*Numbers under lens:* **Evidence-weighted** (`log-odds@1`)
*Machine twins:* https://truthscore.com/claim/metr-found-three-to-six-traces-in-which-an-agent-even-considered--d488cd94.md · JSON at https://truthscore.com/api/v1
METR found three to six traces in which an agent even considered notifying a human.

## Voices

*Counted, never weighed.*

- [reddit.com](https://truthscore.com/entity/9758c1c4) — independent

## Sources

- [OpenAI ExploitGym agents, an Artifactory board, and a Hugging Face breach the graders never scoped](https://truthscore.com/source/e62d3f66)