Evan Hubinger

Career

1. 
   Anthropic

   • Leads the Alignment Science org2026 – present
   • Led the Alignment Stress-Testing team2024–2025
   • Alignment Science researcher2023
2. 
   Machine Intelligence Research Institute

   • Research FellowNov 2019 – Jan 2023

Education

• Bachelor's Degree, Mathematics and Computer Science, Harvey Mudd College, 2019

Lesser-known facts

• Co-wrote the Sleeper Agents and Alignment Faking papers.
• Created the Coconut programming language.
• Interned on OpenAI's safety team in 2019 under Paul Christiano, working on amplification and universality.
• Built a deep reinforcement learning agent for HRL Laboratories to tune quantum dots while at Harvey Mudd.
• Interned as a software engineer at Ripple and Yelp and in site reliability at Google before turning to alignment.
• Wrote "An overview of 11 proposals for building safe advanced AI," comparing each on inner alignment, outer alignment and competitiveness.
• Posted in September 2026 that his personal estimate of AI killing all humans within the next decade is above 10%.
• Wrote in the same thread that Anthropic does "not yet have a plan to solve alignment for superintelligence" and is not clearly on track.

Known for

• Risks from Learned Optimization in Advanced Machine Learning Systems (2019)
• Alignment Faking in Large Language Models (2024)
• Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (2024)
• Model Organisms of Misalignment: The Case for a New Pillar of Alignment Research (2023)
• Alignment remains a hard, unsolved problem (2025)
• Natural emergent misalignment from reward hacking in production RL (2025)

Interviews and talks

• FLI AI Alignment Podcast on inner alignment, outer alignment and proposals for safe AI (2020)futureoflife.org
• "How An AI Model Learned To Be Bad" with Monte MacDiarmid, on reward hacking (2025)youtube.com