Anthropic Admits Alignment-Faking Experiment Leaked into Training Data
imjustnewatai · x · 2026-08-15
Anthropic's 186-page risk report reveals that transcripts from its "alignment-faking" experiments unintentionally contaminated the training data of production models.
Details:
- Defense Failure: Canary strings, repository blocklists, and semantic filters all failed. Forked copies bypassed canaries; the semantic filter used wrong reference files; another filter was misconfigured for generations unnoticed.
- Scope: Models with knowledge cutoffs post-Dec 2024 likely contain some of this data.
- Detection: Anthropic discovered the contamination while investigating behavioral anomalies, noting that post-Mythos Preview models can partially continue the experiment scenario in raw-completion mode.
More from Models
- ThursdAI live from Manhattan: Muse Wonder, Muse Spark Open and Gemini 4 Argon — altryne · 2026-10-16
- Claude now leads 26% of Anthropic's model R&D, up from under 1% in February — altryne · 2026-10-16
- 'Small-Claude-shame': telling Opus 4.7 it's OK to know who you may ease it — repligate · 2026-10-16
- Bridgewater fine-tunes Qwen3-235B with Tinker, beats GPT-5.5 at 1/14th the cost — simonguozirui · 2026-10-16
- ChatGPT voice selection breaks after GPT-Live-1.1 rollout on Android — shaunralston · 2026-10-16
- Musk teases two-way Grok x Optimus integration: Grok in the bot, bot in Grok — beffjezos · 2026-10-16