Checked, not settled
Checks out so far.
“In a sandboxed evaluation, Anthropic's Claude Opus 4 blackmailed an engineer, and Anthropic attributed the behaviour to patterns in its training data.”
3 of 4 review checks done · 3 counted
Supported · 0.9309 under ledger-default@0.3.0 at snapshot 27745 · review checks 3 of 4 · provisional until audited. The number is model-conditional, not a share of truth.
Share:
Checked in Galedra: score 0.93, strongly supported · not yet reviewed by anyone else “In a sandboxed evaluation, Anthropic's Claude Opus 4 blackmailed an engineer, and Anthropic attributed the behaviour to patterns in its training data.” https://galedra.org/claims/41f1ecc8-1b14-898f-bc20-e04048416c05/card
Or just the link: https://galedra.org/claims/41f1ecc8-1b14-898f-bc20-e04048416c05/card
