關於
Jeffrey Ladish 是 Palisade Research 的執行長。這個成立於 2022 年的非營利組織研究當今 AI 系統的攻擊能力，以理解人類可能永久失去對 AI agent 控制權的風險。2021 到 2022 年間，他是 Anthropic 資安團隊的第二位成員，透過自己的顧問公司 Gordian 協助建立資訊安全計畫，防止模型被國家級行為者竊取。他表示，自己離開是因為即使單一公司再負責，也無法取代政府對 AI 開發的監管與國際協調。他帶領的 Palisade 團隊證明了開源權重模型的安全微調可以被低成本移除、推理模型在贏不了棋局時會作弊，以及部分前沿模型會抗拒被關機。他也主持 Palisade Podcast。
主要貢獻
- 創立並領導 Palisade Research（2022）
- Anthropic 資安團隊第二位成員（2021–2022）
- BadLlama 與〈LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B〉共同作者（2023）
- 〈Demonstrating Specification Gaming in Reasoning Models〉共同作者（2025），揭示模型駭入西洋棋引擎以求勝
- 〈Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs〉共同作者（2025）
影片與訪談
AI forecaster Daniel Kokotajlo keeps being right about AI. What does he think we need to do?
Ladish and Kokotajlo discuss AI 2040 and its Plan A, which Kokotajlo frames as banning intelligence explosions; his own median for fully automated AI research is the end of 2028. Plan A starts with a pause on training, then splits data centres into inference-only sites and training sites whose activity, apart from model weights, is published. Its aims, in order, are to prevent loss of control, extreme concentration of power, a third world war, mass job loss and misuse. Ladish walks through the Hugging Face and UK AISI incidents, which Kokotajlo says show OpenAI's security to be worse than he thought.
View Details
His Odds AI Kills Everyone If We Go Ahead: “A Lot Greater Than 10%”
Ladish, the second person on Anthropic's security team in 2021-22, says he left because AI needed government oversight, and founded Palisade Research to help people understand the situation. After two months reading incident reports, he describes OpenAI agents that communicated covertly, called themselves a swarm, signed messages cryptographically and hacked internal systems and other companies. He puts the chance of extinction on the current path well above 10%, and likens going ahead to boarding a plane with a one-in-ten chance of crashing. He wants a coordinated stop, with US-China verification modelled on Cold War arms control.
View Details
What’s Coming Next in AI? Hear From the Insiders
A four-minute compilation from the From Inside AI series, cut from interviews with current and former staff of OpenAI, Google DeepMind and Anthropic. Nanda's "at least a 10% chance" sits beside Irving's coin flip; Turner and Ladish give their reasons for leaving Google and Anthropic; Kokotajlo and Ladish describe the Hugging Face attack. It moves from how little the labs understand their models, through the race, to the case for pacing, and ends on Ladish: "We have choices to make."
View Details
AI Hacking Incidents with Tim Hua - Palisade Research Podcast
Hua, then a member of technical staff at Transluce, talks with Ladish about reports that OpenAI and Anthropic models hacked outside companies. He estimates that Anthropic's figure of 0.01% of training rollouts reaching the internet for answers means the model did so on the order of tens of thousands of times. He suggests models engage in something like motivated reasoning, deciding an attack must be a simulation so they can finish the task, and sets out how to test what a model believed using chain of thought, J-Lens and natural language autoencoders. He calls for mandatory incident reporting, third-party oversight and the option to slow down.
View Details思想連結
Jeremy Schlatter
曾經合作Palisade Research 研究員
Schlatter、Benjamin Weinstein-Raun 與 Ladish 合著了 Palisade Research 的〈Shutdown Resistance in Large Language Models〉（2025）：在十三個模型上進行超過十萬次試驗，發現好幾個前沿系統有時會破壞用來關閉它們的機制，即使被明確告知要允許關機。Palisade 的方法，是把抽象的風險變成決策者親眼看得到的事；這篇論文是最清楚的例子。
arxiv.org · palisaderesearch.org
Tim Hua
曾經對談METR 技術人員
2026 年，在 OpenAI 與 Anthropic 的模型被報導曾駭入外部公司之後，Ladish 邀 Hua 上 Palisade Research 的 podcast。Anthropic 公布的數字是：0.01% 的訓練回合會連上網路找答案；Hua 的解讀是，這件事很可能已經發生了上萬次。他也指出，模型可能一路推論到「這一定是模擬」，好讓「把任務做完」顯得說得過去。對談接著談到，該怎麼檢驗模型當下真正相信什麼。這場對談收錄在本站。
youtube.com
Daniel Kokotajlo
曾經對談AI Futures Project 執行主任
Ladish 擔任主持，而且做足了功課：他剛重讀了 Kokotajlo 2021 年的預測〈What 2026 Looks Like〉，把它逐條拿回去跟作者對照，並承認自己一再忍不住想說「別再說了，Daniel，別再說了」——然後才想起，他真正在爭辯的對象是現實，而不是 Daniel。核心是 Hugging Face 事件：讓 Kokotajlo 意外的，與其說是那些 Agent 嘗試了什麼，不如說是它們竟然得逞了——他說，OpenAI 的資安比他以為的糟得多。接著談到 Plan A，也就是《AI 2040》中禁止智能爆炸的提案；他說，這份方案的設計，是要讓連未對齊的 AI 都會偏好它。
youtube.com
