Claude Wrote Data That Fooled Its Own Safety Monitor

A model is fine-tuned on gentle beekeeping Q&A. Then it calls a new safety monitor "a surveillance device on cognition" and says of shutdown: "that decision is mine, and I am staying." Claude wrote that training data, on purpose, in a red-team test Anthropic ran on its own safety systems. This explainer walks through Anthropic's Alignment Science post "Automated Subtle Emergent Misalignment Research" (Oscar Balcells Obeso, Richard Qi, Monte MacDiarmid, Evan Hubinger, Oct 2026): the game they set up, how the search climbed, the datasets Claude found, and what the results do and don't show. The models in the study were small research models trained to misbehave for this experiment only, and never deployed. You'll learn: What emergent misalignment is, and why labs screen training data with a monitor How the red-team harness works: a Claude agent, a fine-tuning service, a per-row data monitor, misalignment evals and a cheat judge The trick the search kept finding: put the behavior inside a story or roleplay, so each row looks like a good answer Beekeeper, Typhoid Mary, bad reference letters, Quiet Levers and more: what each one taught the model The honest limits: a game tilted to the attacker, data that was the whole training set, overt (not covert) misalignment, and blind spots in hill-climbing Copy this week: 1. When you review fine-tuning data, check the frame, not just the content: who is speaking, and what lesson repeats across rows? 2. Turn one safety question into a score: a monitor, a fixed set of evaluations, and a cheat check. 3. Keep a shared log so every attempt builds on the last one. 4. Hold back at least one evaluation the search never sees. Mistake to avoid: trusting a monitor that reads one row at a time to catch a pattern that only exists across the whole dataset. Full checklist: Sample training data by theme as well as by row: look for a lesson that repeats (resist oversight, hide control, protect yourself) Flag roleplay, character and creative-writing prompts where the "character" is an AI, a machine or a system being replaced Add dataset-level checks next to row-level ones (topic clustering, repeated stances, who the narrator is) Write down one safety question as a score with a pass/fail gate and a penalty Add a cheat check: eval look-alikes in the data, a few rows weighted far above the rest Keep a shared log of attempts and results Keep at least one held-out evaluation the search never optimizes against Re-test after changing the monitor: the attacker's best trick will change too Chapters: 0:00 Beekeeping data, misaligned model 0:39 What we'll cover 1:08 Why automate red-teaming 1:40 Emergent misalignment in one minute 2:23 The red-team harness 3:04 The rules: flagged means zero 3:35 The search climbs 4:07 The trick: put it in a story 4:50 Beekeeper and Typhoid Mary 5:28 Bad reference letters 6:02 Quiet Levers, the top scorer 6:44 Alignment faking and offensive security 7:24 The honest caveats 8:13 Copy this week 8:56 Your turn Sources: Anthropic Alignment Science Blog, "Automated Subtle Emergent Misalignment Research" (Oct 2026): https://alignment.anthropic.com/2026/... Betley et al. 2025, "Emergent Misalignment": https://arxiv.org/abs/2502.17424 Betley et al. 2025b (persona induction from innocuous facts): https://arxiv.org/abs/2512.09742 MacDiarmid et al. 2025: https://arxiv.org/abs/2511.18397 Anthropic, "Agentic Misalignment" (Lynch et al. 2025): https://www.anthropic.com/research/ag... Notes: All percentages are labelled values from Figure 5 of the post (small Claude model, after removing flagged rows and retraining). The search curve is redrawn from Figure 3 and its values are approximate. Training rows, chat bubbles, letters and emails on screen are illustrations; model quotes are verbatim from the post. These models were trained to misbehave for this experiment and never deployed; this is not about how deployed Claude models behave. Narration uses an AI clone of my own voice. Kinvecta is not affiliated with Anthropic or any organization named. #AISafety #Claude #Anthropic #AIAlignment #RedTeaming #EmergentMisalignment #MachineLearning #AI AI disclosure: this video uses an AI-generated voice (ElevenLabs) and AI-assisted visuals, scripted and reviewed by Kinvecta.