Anthropic's account of the changes to Claude's safety training after Claude 4, using its agentic-misalignment honeypots (blackmail, cancer-research sabotage, framing a colleague for financial crimes) as the case study. Experiments fine-tune Claude Sonnet 4, Claude Haiku 4.5 or their base models with synthetic document fine-tuning (SDF) and SFT; RL runs use a haiku-class model and the Sonnet 4 base. Three interventions are compared: SFT on synthetic honeypot transcripts, with and without responses rewritten to reason about ethics; SFT on a 3M-token chat dataset in which the user, not the AI, faces an ethical dilemma (the difficult advice set); and SDF on constitution documents and fictional stories of aligned AIs, which the authors say should "update the distribution of personas that the base model represents". The authors list as limitations that the evals cover only specific scenarios, that experiments are on Sonnet- and Haiku-class models only, that mechanisms are not understood, and that results may depend on Anthropic's own infrastructure. No weights, datasets or eval code are released.
Two versions were published the same day. This stub is for the Alignment Science writeup (bylined, with Methods, ablations and appendix), which the finding treats as primary. The shorter, unbylined Anthropic blog post at https://www.anthropic.com/research/teaching-claude-why carries two things the writeup does not. It says every model since Haiku 4.5 scores 0 on the agentic-misalignment eval, against up to 96% for Opus 4. Its footnote 2 lists Sonnet 4.5 as under 1% but not 0, and says recent models' results may be confounded by information about the eval in pretraining data. The blog does not name the scenario variant behind the 96%. The earlier agentic-misalignment case study gives Opus 4 a 96% blackmail rate in the text-based experiment closest to its computer-use demo (cached at cache/posts/source-2025-agentic-misalignment-lynch.md; that case study has no stub).
Figures: most quantities are plotted without value labels. Numbers cited from this source are those stated in body text or captions (22% → 15% → ~3% → ~1%, 3M vs ~85M tokens, the 2% and 19% ablations, 65% → 19%, over 60% → 25% at 300M tokens, 1.3x–3x). Two values are printed on figures: the Sonnet 4 baseline of 0.22 in the legend of fig7.png, and the best-fit slope of −0.09 per 100M tokens in fig16.png. Values in the name-manipulation chart (fig2.png), the 14M-token stories chart (fig3.png), the auditing-category bars (fig5.png), the RL curves (fig1.png, fig12.png, fig13.png), the self-vs-Claude belief chart (fig11.png) and the RL-mix curves (fig14.png) are axis-only and are not cited. Local copies: cache/posts/source-2026-teaching-claude-why.{html,md} (writeup), cache/posts/source-2026-teaching-claude-why-blog.{html,md} (blog), figures under cache/posts/figures/2026-teaching-claude-why/.