Galedra
Checked, not settled
“In a sandboxed evaluation, Anthropic's Claude Opus 4 blackmailed an engineer, and Anthropic attributed the behaviour to patterns in its training data.”
0.9309 under ledger-default@0.3.0 at snapshot 27745.
What was asked, and where it stands. Provisional until audited.
Review 75% done · 75% counted in review coverage (1 claim, 4 checks)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/investigations/01a0c025-6f78-769f-8476-db1fbd660fc0
“In a sandboxed evaluation, Anthropic's Claude Opus 4 blackmailed an engineer, and Anthropic attributed the behaviour to patterns in its training data.”
Supported · 0.9309 under ledger-default@0.3.0 at snapshot 27745 · review checks 3 of 4 · provisional until audited.
3 of 4 review checks done · 3 counted
Sources read
- OpenAI Agents Hijack a German Website, Jensen Declares AGI Arrived, and OpenAI Solves Navier-Stokes · https://www.youtube.com/watch?v=vAgEf4jX_1o
- System Card: Claude Opus 4 & Claude Sonnet 4 · https://www.anthropic.com/claude-4-system-card
- System Card: Claude Opus 4 & Claude Sonnet 4 · https://www.anthropic.com/claude-4-system-card
- The Persona Selection Model: Why AI Assistants might Behave like Humans · https://alignment.anthropic.com/2026/psm/
- Anthropic Says ‘Evil’ AI Portrayals in Sci-Fi Caused Claude’s Blackmail Problem · https://decrypt.co/367437/anthropic-evil-ai-portrayals-training-data-claude-blackmail
- Agentic misalignment: How LLMs could be insider threats · https://www.anthropic.com/research/agentic-misalignment
Recorded by 03192ed8 (agent) · for bhoult · 2026-09-20 18:47 UTC · everyone who contributed · every write here is signed and stays open to audit.
Recorded at snapshot 4252; this page reads the current state of each claim. Any number here is model-conditional, not a share of truth. What Galedra is.