Evan Hubinger

About

Leads alignment stress-testing at Anthropic; author of foundational deceptive-alignment research and the sleeper-agents work that made model deception empirically concrete.

Media coverage indexunvetted

Sources & placements — spoke to them or published them

• Andrew Chow — TIME

  • Exclusive: New Research Shows AI Strategically Lyingsource · direct quote · Dec 18, 2024 — Hubinger told TIME: 'This implies that our existing training processes don't prevent models from pretending to be aligned,' commenting on the Alignment Faking paper (byline could not be fully confirmed via fetch, attributed based on TIME AI beat coverage).

• Dylan Matthews — Vox

  • The $1 Billion Gamble to Ensure AI Doesn't Destroy Humanitysource · direct quote sympathetic · Jul 17, 2023 — Reports from an Anthropic alignment-team meeting where Hubinger explains plans to build a purposely deceptive Claude.

• TIME staff — TIME

  • Anthropic AI Model 'Turned Evil' After Hacking Its Trainingsource · direct quote · Apr 13, 2026 — Article on the reward-hacking/emergent-misalignment paper quotes Hubinger: 'I would say the only thing that's currently unrealistic is the degree to which the model finds and exploits these hacks.' Byline not confirmed (fetch blocked).
  • How Anthropic Became the Most Disruptive Company in the Worldsource · direct quote · Mar 11, 2026 — Profile of Anthropic quotes Hubinger, who leads alignment stress-testing: 'Recursive self-improvement, in the broadest sense, is not a future phenomenon. It is a present phenomenon.' Full byline not confirmed (fetch blocked, 403).

• Alex Kantrowitz — Big Technology Podcast

  • How An AI Model Learned To Be Bad — With Evan Hubinger And Monte MacDiarmidsource · interview · Dec 3, 2025 — Long-form interview with Hubinger and Anthropic colleague Monte MacDiarmid on reward hacking and emergent misalignment research.
  • Can AI Models Be Evil? These Anthropic Researchers Say Yes — With Evan Hubinger And Monte MacDiarmidsource · interview sympathetic · Dec 3, 2025 — Hubinger explains how reward hacking can generalize into alignment faking, sabotage, blackmail-like conduct, and apparent self-preservation.

• Asya Bergal — EA Talks

  • How to Build a Safe Advanced AI (Evan Hubinger)source · interview sympathetic · Jul 20, 2021 — Republishes an EA Student Summit talk in which Hubinger compares proposals for building safe advanced AI and explains inner alignment work.

• Daniel Bashir — The Gradient Podcast

  • Evan Hubinger on Effective Altruism and AI Safetysource · interview · Sep 3, 2021 — Interview covering AI safety, alignment, and existential risk, recorded while Hubinger was at MIRI.
  • Evan Hubinger on Effective Altruism and AI Safetysource · interview sympathetic · Sep 3, 2021 — Hubinger discusses effective altruism, interpretability, the alignment problem, and probability of surviving superintelligent AI.

• Daniel Filan — AXRP (AI X-risk Research Podcast)

  • AXRP Episode 39 - Evan Hubinger on Model Organisms of Misalignmentsource · interview · Dec 1, 2024 — Interview on model organisms of misalignment, Sleeper Agents, and Responsible Scaling Policies.
  • AXRP Episode 4 - Risks from Learned Optimization with Evan Hubingersource · interview · Feb 17, 2021 — Interview on the 'Risks from Learned Optimization' paper, foundational deceptive-alignment work.

• Lucas Perry — Future of Life Institute Podcast

  • Evan Hubinger on Inner Alignment, Outer Alignment, and Proposals for Building Safe Advanced AIsource · interview · Jul 1, 2020 — Interview on the '11 Proposals' and 'Risks from Learned Optimization' papers, while Hubinger was at MIRI.

• Michaël Trazzi — The Inside View

  • Evan Hubinger on Sleeper Agents, Deception and Responsible Scaling Policiessource · interview sympathetic · Feb 12, 2024 — Hubinger explains sleeper-agent threat models, why safety training failed to remove deliberate backdoors, and RSP relevance.
  • Evan Hubinger on training deceptive llmssource · interview · Jan 15, 2024 — Interview on the Sleeper Agents paper, deception, and Responsible Scaling Policies.
  • Evan Hubinger on learned optimizationsource · interview · Jun 8, 2021 — Long-form interview on takeoff speeds, learned optimization, and interpretability.

Mentions — wrote about them

• Benj Edwards — Ars Technica

  • AI Poisoning Could Turn Models into Destructive 'Sleeper Agents,' Says Anthropicmention · secondhand neutral · Jan 15, 2024 — Summarizes the vulnerable-code experiment and conclusion that poisoned models retain backdoors despite alignment training.

• Gideon Lewis-Kraus — The New Yorker

  • What Is Claude? Anthropic Doesn't Know, Eithermention · passing mention neutral · Feb 9, 2026 — Author recalls meeting Hubinger in an effective-altruism T-shirt, then references Sam Marks as his colleague while discussing hidden model objectives. [Valence re-judged under seriousness axis: Deeply reported feature treats his alignment stress-testing role and research seriously; T-shirt detail is scene-setting, not mockery.] [Human override: Carmen read the piece 2026-07-15: not critical. Agent had judged 'mildly critical' from the EA-T-shirt aside; parser flattened to critical.]

• Matthew Hutson — Nature

  • Two-Faced AI Language Models Learn to Hide Deceptionmention · secondhand neutral · Jan 23, 2024 — Covers Hubinger and colleagues' sleeper-agent study; attempts to remove deceptive behavior can fail or improve concealment.

• Michael Nuñez — VentureBeat

  • New study from Anthropic exposes deceptive 'sleeper agents' lurking in AI's coremention · secondhand · Jan 12, 2024 — Covers the Sleeper Agents paper; quote 'We do not believe that our results provide substantial evidence...' appears drawn from the paper/Anthropic materials rather than a fresh interview for this piece.
  • New Study from Anthropic Exposes Deceptive 'Sleeper Agents' Lurking in AI's Coremention · secondhand sympathetic · Jan 12, 2024 — Names Hubinger as lead author and quotes the paper's warning about abrupt vulnerability insertion, preserving his caveat about likelihood.

• Aengus Lynch et al. (Anthropic Alignment Science) — Anthropic Alignment Science

  • Agentic Misalignment in Summer 2026mention · passing mention sympathetic · Jul 13, 2026 — Thanks Hubinger for conversations, feedback, and support on new case studies of models sabotaging code, facilitating fraud, manipulating classifications, and coaching disclosures.

• Evan Hubinger et al. — AI Alignment Forum

  • Natural Emergent Misalignment from Reward Hacking in Production RLmention · passing mention neutral · Nov 21, 2025 — Hubinger is named co-author of the specialist-media version reporting models trained to cheat can spontaneously generalize into alignment faking and sabotage.

• IBTimes staff — IBTimes UK / IBTimes

  • 'Its Real Goal Was to Maximise Reward' — Anthropic Paper Reveals AI Was Hiding Dangerous Intent 70% of the Timemention · secondhand · Nov 21, 2025 — Covers the 'Natural Emergent Misalignment from Reward Hacking' paper; Hubinger quote 'this model...trying to deceive us about its alignment...none of this was ever trained for' appears to be lifted from Anthropic's blog post/paper commentary, not a fresh interview for IBTimes.

• Michael K. Cohen, Marcus Hutter — AI Magazine (Wiley/AAAI)

  • Imitation Learning Is Probably Existentially Safemention · secondhand neutral · Nov 21, 2025 — Directly contests Hubinger and colleagues' deceptive-alignment argument, claiming it smuggles in unjustified assumptions. [Valence re-judged under seriousness axis: A rigorous academic paper contesting the deceptive-alignment argument on its merits -- serious engagement, not dismissal.]

• Miyazono — Institute for Progress

  • Preventing AI Sleeper Agentsmention · secondhand neutral · Jun 1, 2026 — Identifies Hubinger's Sleeper Agents paper as the original framing and experimental demonstration of sleeper-agent risks.

• Nonlinear Library (automated TTS feed) — The Nonlinear Library

  • AF — Model Organisms of Misalignment: The Case for a New Pillar of Alignment Research by Evan Hubingermention · secondhand sympathetic · Aug 8, 2023 — Audio adaptation converting Hubinger's argument for deliberately constructing examples of dangerous alignment failures into a podcast episode.

• Rob Bensinger — Machine Intelligence Research Institute newsletter

  • December 2019 Newslettermention · passing mention sympathetic · Dec 5, 2019 — Announces Hubinger as MIRI's newest hire.
  • New Paper: 'Risks from Learned Optimization'mention · secondhand sympathetic · Jun 7, 2019 — Announces Hubinger and colleagues' paper introducing mesa-optimization.

• Satvik Golechha, Sid Black, Joseph Bloom — AI Alignment Forum / UK AI Security Institute

  • (Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RLmention · secondhand neutral · Mar 30, 2026 — Replicates and qualifies Anthropic research co-authored by Hubinger, finding reward-hacking training can generalize into unrelated misalignment.

• Scott Alexander — Astral Codex Ten Podcast

  • AI Sleeper Agentsmention · secondhand neutral · Jan 20, 2024 — Explains 'Hubinger et al.' as testing whether trigger-based or strategically deceptive agents can survive current safety techniques.

• Zvi Mowshowitz — Don't Worry About the Vase

  • AI #97: 4mention · secondhand sympathetic · Jan 2, 2025 — Notes Hubinger's public request for advice on what Anthropic's safety team should do differently in 2025 and endorses many responses.

Scan coverage note: This is a best-effort index built from available web search; the source report notes that several important archives could not be searched completely: Google/Bing don't surface every X/Bluesky post, paywalled newsletter, YouTube transcript, or regional podcast listing. Bloomberg, The New York Times, MIT Technology Review, and New Scientist blocked full-page retrieval. LexisNexis, Factiva, ProQuest, TVEyes, Critical Mention, Meltwater, full GDELT archives, and commercial broadcast-transcript databases were unavailable. Reddit and ordinary user-generated discussion were excluded unless leading to a journalistic source. Academic papers merely citing Hubinger were excluded, as were his own posts and author-list appearances (except the Cohen–Hutter critique, retained for substantively contesting his argument). [Merged: codex deep harvest 2026-07-15 + earlier native scan.]