Crosspost from my Substack, One Thousand Means, which covers AI governance from the left. Epistemic status: this is persuasive writing aimed at left-organizing audiences, rather than a neutral policy analysis. I'm posting it here because I think the AI Safety movement underrates the organized left as a coalition partner for...
Main Finding: When you provide classic alignment-faking evaluation prompts to Claude 4.5 Haiku, you see robust non-compliance mediated by jailbreak detection. When you strip adversarial cues from these prompts, Haiku's compliance increases dramatically and shifts from jailbreak refusal to value-based reasoning. Full write-up for this research sprint with figures: alexeigannon.com/adversarialdetection...
> Note: This is a static version of an interactive primer with animated visualizations and a glossary sidebar. I recommend reading the full version, but the text stands on its own. Motivation & Background On the one hand, I write this as a neuroscientist who believes the techniques developed by...