Center for a New American Security (CNAS)

Theory of Change

CNAS's AI safety theory of change is to translate catastrophic AI risk concerns into the language and institutions of US national security, thereby reaching policymakers that the EA/AI safety community cannot directly access. The AI Security and Stability program (launched 2022, funded by Coefficient Giving/Open Phil) articulates five lines of effort: (1) mitigate risks from advanced AI capabilities including misalignment and loss of control; (2) shape compute governance policy; (3) improve US military AI safety and testing; (4) understand Chinese AI decision-making and stability risks; (5) inform US economic security tools for AI diffusion.

The operating mechanism: produce policy reports, convene government and industry leaders, place alumni in government positions, and build a pipeline of national security professionals who understand AI risks. Paul Scharre, EVP and intellectual leader, articulated the core framing on 80K Hours (Dec 2025): "AI technology is powerful but brittle." In private, his views go further -- he called it "bonkers" not to believe AGI could massively transform society and noted his own timelines have been "pulled forward."

The most recent signal of CNAS's direction is "Off Target" (March 24, 2026), which engages directly with technical alignment concepts (deceptive alignment, sleeper agents, reward hacking) and argues alignment "may become the binding constraint" on military AI adoption. This paper cites Anthropic's alignment research extensively and references the Anthropic-DoD dispute of early 2026.

What They Do

Research output. Steady stream of AI-focused reports since 2022: "AI and International Stability" (2023, confidence-building measures between adversaries), "Secure Governable Chips" (2024, on-chip governance mechanisms, co-authored with IAPS), "Catalyzing Crisis" (2024, catastrophic AI risk primer), "Tipping the Scales" (2025, AI and cyber offense-defense balance), "Prepared Not Paralyzed" (2025, AI risk management), chip smuggling analysis, and compute governance papers. "Off Target" (March 2026) is the first paper to engage directly with technical alignment.

Government pipeline. 16+ CNAS experts and alumni entered the Biden administration in January 2021, including Victoria Nuland (Under Secretary of State for Political Affairs), Colin Kahl (Under Secretary for Defense), David Cohen (CIA Deputy Director), Kurt Campbell (Deputy Secretary of State), and Avril Haines (DNI). This revolving door is CNAS's single most powerful asset.

Policy engagement. Responded constructively to both the Biden AI Executive Order and Trump AI Action Plan, demonstrating ability to work across administrations. Submitted detailed recommendations on AI governance, export controls, and military AI safety.

Convening. AI Leadership Forum brings together government, industry, and academic leaders. Wargaming exercises (including Taiwan scenarios) incorporate AI and autonomous systems dynamics.

Team expansion. Dec 2025 hired James Sanders (Epoch AI researcher, former quant at SIG), Daniel Remler (State Dept emerging tech), and Liam Epstein (Cambridge AI safety education). The AI Security and Stability team now has 12+ experts.

Key People

Paul Scharre, EVP. Former Army Ranger (Iraq, Afghanistan). Drafted DoD Directive 3000.09 (first US policy on autonomous weapons). Author of "Army of None" and "Four Battlegrounds." TIME 100 AI 2023. The intellectual architect of CNAS's AI safety work. On 80K Hours (Dec 2025), he took transformative AI seriously: "my expectations for how fast this is unfolding have certainly been pulled forward." He identifies biological weapons and cyber-AI intersection as "the scary stuff" and notes national security professionals dismiss AI escape scenarios as sci-fi even while taking nuclear war games seriously.

Richard Fontaine, CEO. Former McCain foreign policy advisor, NSC, State Dept. Also sits on Anthropic's Long-Term Benefit Trust and National Security Advisory Council.

Michele Flournoy, Chair. Co-founder CNAS. Former Under Secretary of Defense for Policy (Obama). Co-founder WestExec Advisors. Board member Booz Allen Hamilton.

Notable staff: Vivek Chilukuri (director, Tech & Nat Sec, former Senate Intel Committee), Tim Fist (senior adjunct, AI hardware expert, co-authored on-chip governance with IAPS), Caleb Withers (research associate, lead author of "Off Target," former NZ government), Michael Sellitto (adjunct, also Anthropic's Head of Geopolitics).

~50 employees total plus many adjunct fellows. Recent hires signal deliberate connection to AI safety pipeline.

Money and Incentives

Revenue: $14.4M (2023), $18.7M peak (2022). Net assets $23.7M (2023). 14 years of 990 data show revenue ranging from $6M to $19M.

CG/Open Phil funding: $14.6M total across 8 grants (2017-2025). Largest: $8.3M (June 2025, 3-year) and $5.1M (July 2022, 3-year). During active grant periods, CG/Open Phil likely represents 30-50% of annual revenue. This is by far the largest single funding source for CNAS's AI safety work and likely the organization's overall largest funder in recent years.

Defense contractor funding: All Big 5 defense contractors are donors. FY2025: Northrop Grumman ($500K+), Lockheed Martin ($100-249K), Anduril ($100-249K), BAE ($50-99K), Boeing ($25-49K), General Dynamics ($25-49K). CNAS was the single largest recipient of defense contractor money among 50 major US think tanks (2014-2019).

Tech company funding (FY2025): Amazon ($500K+), AWS ($100-249K), Google ($100-249K), Meta ($100-249K), NVIDIA ($50-99K), Scale AI ($50-99K), Microsoft ($50-99K).

Anthropic-adjacent money chain: Amazon ($8B Anthropic investor) gave $500K+ to CNAS. Google ($3B+ Anthropic investor, 14% owner) gave $100-249K. CEO Fontaine sits on Anthropic's LTBT. Anthropic's Head of Geopolitics is a CNAS adjunct.

Foreign government funding: $2.8M total (2019-2023). FY2025: Japan ($250-499K), Taiwan ($100-249K), Germany ($50-99K), Korea ($50-99K), Lithuania ($25-49K). Taiwan funded CNAS while CNAS advised strengthening US-Taiwan security ties.

US government funding (FY2025): DoD CDAO ($500K+), DTRA ($250-499K), DARPA ($100-249K).

EA-adjacent funding beyond CG/Open Phil: Founders Pledge ($100-249K), returned $25K FTX donation.

Business model: Grant-funded nonprofit think tank. Revenue is diversified across defense industry, tech sector, foreign governments, US government, EA funders, and individual donors, but every major funding category has direct interests in CNAS's policy recommendations.

Key incentive tension: CNAS's defense contractor and military funders have financial interests in AI-enabled weapons systems. CG/Open Phil funds CNAS to work on AI safety. These interests overlap (e.g., safe military AI) but can diverge (e.g., autonomous weapons governance, export controls that limit AI deployment).

What Others Say

Foreign-policy critique: "A haven for hawkish Democrats" (LA Times via Militarist Monitor). Andrew Bacevich compared CNAS to AEI, saying both see war as "a perpetual condition." In These Times documented CNAS's record of "pushing Democrats to embrace war and militarism." Flournoy endorsed pre-emptive war doctrine (2002), pushed for Libya intervention, and consistently advocated for increased defense spending.

Conflict of interest documentation: The Revolving Door Project (Feb 2021) documented that CNAS "repeatedly violated its own ethics policy" -- recommending B-21 bomber purchases without disclosing Northrop Grumman funding, praising PMCs while funded by them, and accepting Taiwan and UAE money for studies on US policy toward those countries.

Anthropic conflicts: Politico (Oct 2025) reported that Janet Egan praised Anthropic's chip diversification without disclosing Fontaine's LTBT membership, Sellitto's adjunct role, or Amazon/Google's dual status as Anthropic investors and CNAS donors.

AI safety community view: A LessWrong review of "Four Battlegrounds" criticized Scharre for not engaging with AGI timelines and suggested his China focus "suggests he's mostly focused on convincing the US to go to war with China." However, Scharre's more recent statements (80K podcast, "Off Target" paper) suggest significant evolution in his views.

Defenders: Founders Pledge assessed CNAS as having a "strong reputation for producing useful research." Anthropic CEO Dario Amodei, announcing Fontaine's LTBT appointment: "ensuring democratic nations maintain leadership in responsible AI development is essential for both global security and the common good."

What's Absent

No internal dissent or notable departures with public statements have been found. No CNAS researcher has posted on LessWrong, EA Forum, or Alignment Forum. No evidence exists of how CG/Open Phil funding changed CNAS's trajectory -- whether the AI safety work represents organic evolution or funded adoption of EA priorities. No public response from CNAS to the Revolving Door Project's documented conflicts of interest. No metrics exist for evaluating the actual policy impact of specific CNAS reports.