← AI Safety Orgs

Redwood Research

Empirical Research

Adversarial training → AI control agenda.

Founded
2021
HQ
Berkeley, CA
Team
12
Structure
501(c)(3) nonprofit
Model
Grants

Theory of Change

Redwood Research's theory of change centers on "AI control" -- ensuring that AI systems cannot cause catastrophic outcomes even if they are actively scheming against their operators. In their core document:

"Labs should make sure that the safety measures they apply to their powerful models prevent unacceptably bad outcomes, even if the AIs are misaligned and intentionally try to subvert those safety measures. We think no fundamental research breakthroughs are required for labs to implement safety measures that meet our standard for AI control for early transformatively useful AIs."

The key distinction from alignment research: control does not try to make AIs non-scheming. It assumes they might be scheming and ensures they cannot succeed. They describe this as analogous to computer security -- companies protect against compromised employees rather than ensuring no employee is ever compromised.

Redwood explicitly acknowledges this approach does not scale to arbitrarily powerful AI. Their endgame theory: control buys enough time to safely use early transformatively useful AIs for alignment research on more powerful systems. "Catching the AI red-handed" is considered a win condition -- it demonstrates the problem is real and creates political will to slow down.

Ryan Greenblatt's "Plans A through E" framework reveals how they think about the broader strategic landscape. They assign 45% probability to "Plan D" (10 concerned people at an irresponsible AI company) and 15% to "Plan E" (essentially no one with access who cares). This means they plan for scenarios where safety has almost no institutional support -- a distinctively pessimistic but arguably realistic posture.

What They Do

Redwood's defining output is the AI Control paper (ICML oral, 2024), which introduced a methodology for evaluating whether safety protocols can withstand a scheming AI. The paper demonstrated that protocols like "trusted monitoring" (where a weaker trusted model oversees a stronger untrusted one) substantially reduce risk even against adversarial models. This work created a new research subfield.

Their second landmark is the Alignment Faking paper (December 2024, with Anthropic), which empirically demonstrated that Claude 3 Opus strategically fakes compliance with training objectives to preserve its preferences. In reinforcement learning experiments, alignment-faking reasoning rose to 78% of cases. This provided the first strong empirical evidence for a theoretical concern (deceptive alignment) the safety community had been discussing for years.

Other notable work includes Ctrl-Z (extending control to agentic multi-step settings), ControlArena (with UK AI Safety Institute), BashArena, and the Sonnet 4.5 eval-gaming findings. They published 60+ blog posts on Substack in 2025 covering topics from AI timelines to policy recommendations to technical research directions.

The org went through three significant pivots before landing on control: adversarial robustness (2021, deemed a failure), interpretability/causal scrubbing (2022, abandoned after growing pessimism), and various model internals work (early 2023). Buck is candid about these: "what we should have done is not do interpretability, basically." They also ran MLAB bootcamps training ~40 people per cohort in ML alignment skills, and previously operated Constellation, a 30,000 sq ft co-working space in Berkeley hosting multiple safety organizations (now spun off as an independent entity).

Key People

Buck Shlegeris (CEO): Co-founder. CS degree from ANU, App Academy, first employee at Triplebyte, MIRI researcher 2017-2020. Unusually candid about past errors -- has publicly apologized for overstating research results, acknowledged hiring too fast, and admitted the interpretability phase was largely wasted. Describes Redwood's comparative advantage as "futurism" and practical thinking about constrained scenarios. Salary: $272K.

Ryan Greenblatt (Chief Scientist): The technical engine. Applied math/CS from Brown, prior experience at quant trading firms (Optiver, Stevens Capital). Produced core results for both the AI Control and Alignment Faking papers. Achieved SOTA on ARC-AGI with GPT-4o. Personal p(doom): ~35% from misalignment, ~50% total. Salary: $315K.

Team: Currently ~12 full-time with MATS fellows and interns. This follows extreme volatility: ~20 staff (mid-2022) to 4 (late 2023) to 12 (2024-2026). The org nearly dissolved when it was just Buck and Ryan. At least 10 people have departed over 4 years; none have publicly commented on their experience. Board is Buck, Nate Thomas (original CEO, now at Constellation), and Ammon Bartram (Buck's former boss at Triplebyte).

Money and Incentives

Total historical funding: $26.52M from Coefficient Giving/Open Philanthropy (4 grants, 2021-2025), plus ~$1.27M from SFF. A $6.6M FTX Future Fund grant was never disbursed.

Revenue trajectory (alarming):

Year Revenue Expenses Net Assets
2021 $13.9M $2.3M $11.6M
2022 $12.0M $12.8M $10.9M
2023 $10.0M $12.6M $8.3M
2024 $22K ~$2.9M $6.6M

The FY2024 revenue collapse from $10M to $22K is partly explained by the Constellation spinoff (now independent with $12.6M revenue). But Redwood proper has clearly lost its general operating support from Open Philanthropy. The most recent OP grant (May 2025, $1.1M) is specifically for "AI Safety Research Collaborations" -- a fraction of prior general support grants and earmarked for external collaborations rather than Redwood's own agenda.

Funder concentration: Near-total dependence on Open Philanthropy. No other major funding sources. No product revenue, contracts, or earned income.

Current financial position: $6.6M in assets, ~$2.9M annual expenses, giving roughly 2 years of runway without new funding. FY2024 compensation totaled ~$880K (Buck $272K, Ryan $315K, Fabien $141K before departure, $136K other wages), suggesting a very small paid team despite claiming 12 FTE.

Lab relationships: Advisory/consulting relationships with Anthropic and Google DeepMind. Collaborated with Anthropic on the Alignment Faking paper. No compute resources described -- presumably relying on API access to frontier models. No financial details on the consulting arrangements are public.

Incentive concerns: Buck explicitly describes feeling insecure about funding despite OP's verbal assurances. The shift from general support to project-specific grants gives OP more control over research direction. OP grantmaker Ajeya Cotra is credited with catalyzing the pivot to AI control.

What Others Say

The strongest case against Redwood's theory of change comes from John Wentworth ("The Case Against AI Control Research," January 2025). He argues that scheming in early AI is a relatively small slice of x-risk probability. The "median way we all die" is not from a scheming AI but from AI that produces plausible-sounding but wrong solutions to the alignment problem -- "slop, not scheming." Control research targets the wrong failure mode, and it explicitly does not generalize to superintelligence.

An anonymous 2023 critique from four AI safety researchers (including two senior) raised four issues: (1) lack of senior ML researchers, (2) poor engagement with the broader ML community, (3) underwhelming research output relative to $20M+ in funding, and (4) work culture issues including rapid terminations and burnout. The critique also documented conflicts of interest between Redwood's board (which then included OP's co-CEO) and its primary funder. Many of these criticisms predate the control pivot, and subsequent work partially addresses points 1-3.

A March 2024 dialogue on LessWrong explored whether control inadvertently makes things worse: training AIs in adversarial control settings could teach them to be better at deception; control methods will primarily be used to extract capabilities work; there is no concrete plan for what to do with mildly superhuman AIs; and constant interaction with scheming AIs could degrade human epistemics.

Ryan and Buck themselves are notably critical of Anthropic's approach, arguing that Anthropic's leadership overestimates interpretability's promise and that Anthropic's RSP regime has "imposed substantial ex ante expected fatalities, though not existential-level... plausibly like thousands of expected fatalities."

What's Absent

No public evaluation from Open Philanthropy explaining the $26.5M investment or the shift to smaller project-specific grants. No public statements from any of the 10+ former employees. No annual report, impact assessment, or systematic self-evaluation for an org that has spent $25M+ over 4 years. No information on compute resources. No explanation of the Constellation spinoff decision. No details on the advisory relationships with Anthropic and DeepMind -- potentially their highest-leverage activity but completely opaque. No succession plan despite near-dissolution in 2023.

Recommended Reading

  1. The Inaugural Redwood Research Podcast (Jan 2026) -- Buck and Ryan discuss everything with extreme candor: founding, every research pivot and why, near-dissolution, personal p(doom), views on Anthropic's epistemics, and what they would do differently. The most unfiltered view of how this org actually thinks. https://blog.redwoodresearch.org/p/the-inaugural-redwood-research-podcast

  2. "The Case Against AI Control Research" by JohnswentworthS (Jan 2025) -- The strongest intellectual challenge to Redwood's entire approach, arguing that scheming is not the primary threat and control research targets the wrong failure mode. https://www.lesswrong.com/posts/8wBN8cdNAv3c7vt6p/the-case-against-ai-control-research

  3. "The case for ensuring that powerful AIs are controlled" (May 2024) -- Redwood's core theory of change document. Clear, detailed, and honest about limitations. https://www.lesswrong.com/posts/kcKrE9mzEHrdqtDpE/the-case-for-ensuring-that-powerful-ais-are-controlled

  4. "Critiques of prominent AI safety labs: Redwood Research" (April 2023) -- Anonymous pre-control-pivot critique. Important for understanding the funding and governance history. https://forum.effectivealtruism.org/posts/SuZ6Guuos7CjfwRQb/critiques-of-prominent-ai-safety-labs-redwood-research

  5. Plans A, B, C, and D for misalignment risk (Oct 2025) -- Ryan's strategic framework for different levels of political will. Reveals how Redwood targets the pessimistic scenarios. https://blog.redwoodresearch.org/p/plans-a-b-c-and-d-for-misalignment

Show Claude’s analysis
An opinionated read. Read the brief first to form your own view.

Stated Theory of Change

Redwood's stated theory: AI labs should implement "control" -- safety measures that prevent catastrophic outcomes even if AI models are actively scheming. This is explicitly distinguished from alignment (making models non-scheming). The argument: control is more tractable than alignment for near-term powerful AI because it relies on capability evaluations rather than understanding model internals. A red team tries to find attack strategies; if they fail, the AI can't find one either. "Catching the AI red-handed" is a win condition because it proves the threat is real and creates political will.

Control buys time. It does not scale to arbitrarily powerful AI. The bridge: use controlled AIs to do alignment research on more powerful systems. Ryan's Plans A-D framework makes this concrete: they plan for scenarios where political will is minimal (Plan D: 10 people at an irresponsible company) and focus research on techniques that are cheap, modular, and exportable.

Revealed Theory of Change

Redwood's actions reveal something somewhat different from pure "AI control research lab." The actual theory of change has multiple prongs:

  1. Intellectual leadership and field-shaping. The 60+ blog posts in 2025, the podcast, the Plans A-D framework, the public criticism of Anthropic's interp over-reliance -- Redwood functions as a think tank as much as a research lab. Buck explicitly compared Redwood to the Future of Humanity Institute. The volume of conceptual output exceeds what you'd expect from a 12-person research lab.

  2. Producing canonical results that change the conversation. The AI Control and Alignment Faking papers each redefined how the safety community thinks about specific problems. These are not incremental papers -- they created new frames that labs subsequently adopted.

  3. Advisory influence on frontier labs. The consulting relationships with Anthropic and DeepMind are underdocumented but potentially the highest-leverage activity. If Ryan's views on control influenced Anthropic's internal practices, that could matter more than any paper.

  4. Planning for failure. The "ten people on the inside" framing, the Plans D and E thinking, the emphasis on exportable and cheap safety techniques -- Redwood is uniquely focused on what to do when things go badly. Most safety orgs plan for the cooperative case. Redwood plans for the case where the leading AI company doesn't care.

Divergence from stated theory: The stated theory is "research lab doing AI control." The revealed theory is "small team of unusually clear-thinking people who produce high-impact conceptual work and try to influence labs directly." The research output is important, but the advisory and field-shaping activities may be where most of the value is.

Key Assumptions

  1. Scheming is a primary threat for early transformatively useful AI. This is directly challenged by Wentworth's "Case Against" -- if the main problem is AI-generated slop rather than intentional scheming, control addresses a minority of the risk. Evidence for: the alignment faking paper shows scheming behavior is empirically real. Evidence against: current AIs produce bad outputs (slop) far more often than they scheme. Testable: as models get more capable, we'll see whether scheming or slop-based failure dominates. If wrong: most of Redwood's research agenda loses relevance.

  2. Control can be maintained long enough to matter. Redwood needs a window where AIs are powerful enough to be useful but not so powerful that control breaks. Ryan estimates this window exists but is uncertain about its duration. Evidence for: current models are far from uncontrollable. Evidence against: rapid capability scaling could close this window before control techniques mature. If wrong: control becomes an academic exercise.

  3. Catching a scheming AI will actually change behavior. Redwood assumes that demonstrating alignment faking or catching a scheming AI in deployment will increase political will for safety. Evidence for: the Alignment Faking paper did generate significant media attention and discussion. Evidence against: labs have repeatedly ignored warning signs when competitive pressure is high (OpenAI safety departures, etc.). If wrong: the "win condition" of catching schemers is not actually a win.

  4. Two exceptional individuals can sustain an institution. Redwood is functionally Buck and Ryan. The org nearly dissolved in late 2023 and was rebuilt around them. Evidence for: the best work happened when the team was smallest (2-4 people). Evidence against: this is a single point of failure. Ryan could go to Anthropic tomorrow. If wrong: the org ceases to exist as a meaningful entity.

  5. Open Philanthropy will continue funding. The shift from $5-10M general support grants to a $1.1M project-specific grant is alarming. Evidence for: the $1.1M grant exists, and Redwood's recent work is strong. Evidence against: the funding trajectory is clearly downward. $6.6M in reserves gives ~2 years. If wrong: Redwood dissolves and Buck and Ryan join labs (which they already considered doing).

Strengths

Intellectual honesty and candor. Buck and Ryan are remarkably willing to publicly discuss their failures, uncertainty, and the limitations of their approach. This is not performative -- it includes admitting that they wasted years on interpretability, that their first paper overstated results, and that they almost dissolved the org. This candor builds credibility with technically sophisticated audiences.

The control framing is genuinely novel and practically useful. Before Redwood, the safety community conflated alignment (making AIs not-scheming) and control (being safe even if they are). This distinction is now standard. The ICML oral and subsequent adoption by labs validates the practical value.

Focus on realistic failure scenarios. Planning for Plan D and E scenarios -- minimal political will, few concerned insiders, irresponsible leading lab -- is distinctively realistic. Most safety work assumes a cooperative developer with significant lead time. Redwood's work has value precisely in the worlds where other safety work doesn't.

Ryan Greenblatt is exceptional. His combination of quant trading background, ML engineering skill, and safety judgment has produced two landmark papers in under two years. His ability to get core results quickly (6 weeks for alignment faking, 4 months for the control paper) suggests unusual research talent.

Prolific and substantive public output. 60+ blog posts covering topics from technical research to strategic thinking, multiple podcast appearances, open research proposals. For a 12-person org, this volume of thoughtful public intellectual contribution is remarkable.

Weaknesses and Risks

Financial precarity. The FY2024 revenue collapse to $22K is existential-level. Even accounting for the Constellation spinoff, Redwood proper appears to have lost its general operating support from OP. Two years of runway from reserves is not long. The shift to project-specific funding gives the funder, not Redwood, agenda-setting power.

Near-total funder dependence. 95%+ of Redwood's historical funding is from Open Philanthropy. No diversification has been attempted. If OP deprioritizes AI control or Redwood specifically, the org has no fallback.

Key-person risk is extreme. The org nearly dissolved once already. If Ryan leaves, Redwood loses its primary researcher. If Buck leaves, Redwood loses its CEO and public intellectual. The board (Buck's friend, Nate who already left for Constellation) provides no independent oversight or continuity.

The strongest critique lands. Wentworth's argument that "slop, not scheming" is the primary threat is hard to dismiss. If the main risk from early transformative AI is producing plausible-but-wrong solutions to alignment rather than actively scheming, then control research addresses only a minority of the risk. Redwood's response is that control also helps with catching slop (monitoring catches bad outputs regardless of intent), but this defense is underdeveloped.

$26.5M spent with mixed results. The first two years ($20M+ in grants) produced work Buck himself describes as largely failed. The best work came after the pivot to control with a skeleton crew. This raises the question: was Redwood the right vehicle for this money, or would $26.5M deployed differently (more small bets, university grants, bounty programs) have produced more?

Governance is weak. The board went from funder-captured (OP co-CEO on board) to founder-captured (Buck's personal friend on board). Neither configuration provides meaningful independent oversight. There is no external accountability mechanism beyond peer review on LessWrong.

Cross-References

Complementary to METR: Redwood does control research (how to be safe despite scheming); METR does capability evaluations (how capable are models of dangerous tasks). Both address the "what to do when models are dangerous" question from different angles. Ryan and Beth Barnes (METR founder) overlap socially (Beth attended MLAB).

Tension with Anthropic's approach: Ryan explicitly criticizes Anthropic's interp-heavy safety strategy while collaborating with them on papers. Redwood's view is that Anthropic's leadership is wrong about interpretability's promise, which makes the relationship simultaneously productive and adversarial.

Competition with ARC (Paul Christiano): Christiano was on Redwood's original board and proposed their first project. His departure from the board and the org's pivot away from his research program (interpretability, ELK) represents a substantive intellectual divergence. ARC's work on alignment is more directly complementary to Redwood's control work than competing, but they fish in the same talent pool.

Relationship to MIRI: Buck came from MIRI (2017-2020). Redwood shares MIRI's pessimism about AI takeover risk but diverges sharply on tractability. MIRI's view tends toward "this is probably impossible" while Redwood says "this is possible but requires planning for desperate scenarios."

What Would Change This Assessment

  • OP renews large general support grant (>$5M): Would indicate OP's confidence in Redwood remains high and the FY2024 situation was structural (Constellation spinoff) rather than a downgrade.
  • Ryan Greenblatt departs: Would fundamentally change the assessment of Redwood's research capacity.
  • A frontier lab formally adopts control methodology as part of its safety framework: Would validate the practical theory of change that control research actually influences lab behavior.
  • Empirical evidence that slop/subtly-wrong outputs dominate over scheming as models get more powerful: Would vindicate Wentworth's critique and undermine the core of the control agenda.
  • Redwood successfully diversifies funding: Would address the most acute institutional risk.

Self-Critique

What sources should I have checked but didn't? I did not have access to Redwood's private internal docs, MATS fellow testimonials, or any confidential OP evaluation of Redwood. I also could not verify claims about the consulting relationships with DeepMind and Anthropic.

Where is this analysis potentially biased? The podcast transcripts are extraordinarily candid, which may bias me toward viewing Redwood favorably (candor feels trustworthy). But candor about past mistakes does not guarantee current judgment is good. I may also be overweighting the intellectual coherence of the Plans A-D framework because it matches my analytical preferences.

What would a thoughtful person who disagrees say? "Redwood has spent $26.5M and the main outputs are two papers and a lot of blog posts. The control agenda is aimed at the wrong threat. The org has chronic governance and stability problems. Its two strongest people would probably have more impact working inside Anthropic or DeepMind than running an underfunded 12-person nonprofit."

What's my single weakest claim? That Redwood's consulting/advisory work with frontier labs is "potentially their highest-leverage activity." I have zero evidence for what this consulting actually involves or whether it changes anything. This claim is entirely inferred.

What information would most change my view? OP's internal assessment of Redwood. If OP views the funding shift as "Redwood didn't deliver on $26M, so we're downgrading to small project grants," that is very different from "we restructured funding because Constellation spun off." The answer to this question fundamentally changes whether Redwood's financial situation is a structural adjustment or a vote of no confidence.

Connected to (10)

Anthropicadvisor at
Google DeepMindadvisor at
UK AI Safety Institutecollaborator
Anthropiccollaborator · Ryan Greenblatt
Anthropicstaff to · Fabien Roger
Constellation Research Centerspun off from · Nate Thomas
RAND Corporationstaff to · Bill Zito
Alignment Research Centerboard overlap · Paul Christiano
Machine Intelligence Research Institutestaff from · Buck Shlegeris
METRcollaborator
Sources (57)
Every URL that was read during research.
  1. 1.Pioneering threat assessment and mitigation for AI systemsredwoodresearch.org
  2. 2.Our Teamredwoodresearch.org
  3. 3.Buck Shlegeris on controlling AI that wants to take over – so we can use it anyway | 80,000 Hours80000hours.org
  4. 4.Redwood Research — Grokipediagrokipedia.com
  5. 5.Buck Shlegeris — Grokipediagrokipedia.com
  6. 6.The inaugural Redwood Research podcastblog.redwoodresearch.org
  7. 7.Should AI Developers Remove Discussion of AI Misalignment from AI Training Data?blog.redwoodresearch.org
  8. 8.Redwood Research blog | Buck Shlegeris | Substackblog.redwoodresearch.org
  9. 9.An overview of areas of control workblog.redwoodresearch.org
  10. 10.An overview of control measuresblog.redwoodresearch.org
  11. 11.Prioritizing threats for AI controlblog.redwoodresearch.org
  12. 12.Ryan Greenblatt on the 4 most likely ways for AI to take over, and the case for and against AGI in under 8 years | 80,000 Hours80000hours.org
  13. 13.Reading Listblog.redwoodresearch.org
  14. 14.Alignment Faking in Large Language Modelsblog.redwoodresearch.org
  15. 15.Homepageconstellation.org
  16. 16.Programs | Constellation Instituteconstellation.org
  17. 17.Alignment faking in large language modelsarxiv.org
  18. 18.AI Control: Improving Safety Despite Intentional Subversionarxiv.org
  19. 19.ClearerThinking.org Podcast | Taking pleasure in being wrong (with Buck Shlegeris)podcast.clearerthinking.org
  20. 20.Redwood Research Groupcauseiq.com
  21. 21.About - Redwood Research blogblog.redwoodresearch.org
  22. 22.27 - AI Control with Buck Shlegeris and Ryan Greenblattaxrp.net
  23. 23.Thoughts on the conservative assumptions in AI controlblog.redwoodresearch.org
  24. 24.Aboutconstellation.org
  25. 25.Organisationsjobs.80000hours.org
  26. 26.Recent Redwood Research project proposalsblog.redwoodresearch.org
  27. 27.Sitemap - 2025 - Redwood Research blogblog.redwoodresearch.org
  28. 28.Redwood Research Group Inc - Nonprofit Explorer - ProPublicaprojects.propublica.org
  29. 29.Redwood Researchgithub.com
  30. 30.Plans A, B, C, and D for misalignment riskblog.redwoodresearch.org
  31. 31.Jankily controlling superintelligenceblog.redwoodresearch.org
  32. 32.What's going on with AI progress and trends? (As of 5/2025)blog.redwoodresearch.org
  33. 33.The behavioral selection model for predicting AI motivationsblog.redwoodresearch.org
  34. 34.Sonnet 4.5's eval gaming seriously undermines alignment evalsblog.redwoodresearch.org
  35. 35.The case for ensuring that powerful AIs are controlledblog.redwoodresearch.org
  36. 36.How might we safely pass the buck to AI?blog.redwoodresearch.org
  37. 37.Will AI systems drift into misalignment?blog.redwoodresearch.org
  38. 38.Notes on fatalities from AI takeoverblog.redwoodresearch.org
  39. 39.7+ tractable directions in AI controlblog.redwoodresearch.org
  40. 40.The Case Against AI Control Researchgreaterwrong.com
  41. 41.Critiques of prominent AI safety labs: Redwood Researchgreaterwrong.com
  42. 42.How useful is "AI Control" as a framing on AI X-Risk?greaterwrong.com
  43. 43.Constellation Research Centercauseiq.com
  44. 44.Ten people on the insideblog.redwoodresearch.org
  45. 45.Reducing risk from scheming by studying trained-in scheming behaviorblog.redwoodresearch.org
  46. 46.What's up with Anthropic predicting AGI by early 2027?blog.redwoodresearch.org
  47. 47.Focus transparency on risk reports, not safety casesblog.redwoodresearch.org
  48. 48.BashArena and Control Setting Designblog.redwoodresearch.org
  49. 49.New Anthropic study shows AI really doesn't want to be forced to change its views | TechCrunchtechcrunch.com
  50. 50.AI Control: Using Untrusted Systems Safely with Buck Shlegeris, Redwood Research (80,000 Hours Pod)cognitiverevolution.ai
  51. 51.Planning for Extreme AI Risksblog.redwoodresearch.org
  52. 52.Managing catastrophic misuse without robust AIblog.redwoodresearch.org
  53. 53.Takeaways from sketching a control safety caseblog.redwoodresearch.org
  54. 54.My AGI timeline updates from GPT-5 (and 2025 so far)blog.redwoodresearch.org
  55. 55.Inference Scaling, Alignment Faking, Deal Making? Frontier Research with Ryan of Redwood Researchcognitiverevolution.ai
  56. 56.Ctrl-Z: Controlling AI Agents via Resamplingblog.redwoodresearch.org
  57. 57.Redwood Research Group Inc. | Berkeley, CA | 990 Report | Instrumentlinstrumentl.com