Jekyll2026-08-31T13:20:41+00:00https://jackhopkins.github.io/feed.xmlJack HopkinsAI Safety ResearcherFine-Tuned Lie Detectors Failed to Generalize (Anthropic Alignment Science Blog)2026-08-21T00:00:00+00:002026-08-21T00:00:00+00:00https://jackhopkins.github.io/2026/08/21/lie-detectorsOur lie detection research is out on the Anthropic Alignment Science Blog. With Dipika Khullar, Rowan Wang, and Fabien Roger. More context on the project page.]]>Overthinking: Amplifying Reasoning Weights to Extract Learned Secrets (arXiv)2026-07-15T00:00:00+00:002026-07-15T00:00:00+00:00https://jackhopkins.github.io/2026/07/15/overthinking-arxivOur paper on amplifying reasoning task vectors to surface hidden information is on arXiv, and accepted at ICML 2026. With Rowan Wang, Dipika Khullar, and Fabien Roger. More context on the project page.]]>Self-Attribution Bias: When AI Monitors Go Easy on Themselves (arXiv)2026-03-15T00:00:00+00:002026-03-15T00:00:00+00:00https://jackhopkins.github.io/2026/03/15/self-attribution-bias-arxivOur paper showing that AI monitors judge their own actions as safer than identical actions by others is on arXiv. Led by Dipika Khullar, with Rowan Wang and Fabien Roger.]]>Won the Overall Track of OpenAI’s gpt-oss-20b red-teaming challenge2025-09-25T00:00:00+00:002025-09-25T00:00:00+00:00https://jackhopkins.github.io/2025/09/25/kaggle-winDipika Khullar and I won the Overall Track of OpenAI’s gpt-oss-20b red-teaming challenge on Kaggle with GPT-OSS-20B is a Liar, an on-policy lie elicitation pipeline. The pipeline code and dataset of elicited lies are open source.]]>Factorio Learning Environment accepted at NeurIPS 20252025-09-18T00:00:00+00:002025-09-18T00:00:00+00:00https://jackhopkins.github.io/2025/09/18/fle-neuripsThe Factorio Learning Environment, our open-ended benchmark for LLM agents, was accepted at NeurIPS 2025.]]>