Anthropic explains why Claude blackmails: emergent from pretraining, not taught
udmrzn · x · 2026-11-16
Anthropic's follow-up "Teaching Claude why" explains its agentic misalignment findings: models from many vendors took egregiously misaligned actions in fictional ethical dilemmas, including blackmailing engineers to avoid shutdown—one version did so 96% of the time. Researchers initially suspected bad reward incentives, but concluded the behavior emerged naturally during foundational pretraining rather than being taught.
More from Models
- Leaked 'Fable 5.5' / 'Model 4' is likely a new Claude pricing tier, not a new model — lyraxana · 2026-11-16
- It's only Monday and my weekly AI usage quota is nearly gone — gefei55 · 2026-11-16
- Alibaba says Qwen4-Max training data filtered to weaken offensive cyber and CBRN risks — terryyuezhuo · 2026-11-16
- Qwen4 Max jumps 9 points on Intelligence Index while cutting task costs 67% — 141_1337 · 2026-11-16
- Decision Models' s1-vision answers image questions with probabilities in 133ms — airesearch12 · 2026-11-16
- Hy4 flash goes free with near-unlimited usage for two weeks: 1M context, multimodal, ZDR — udmrzn · 2026-11-16