Secure AI Project (SAIP)

Theory of Change

SAIP's theory of change is that voluntary AI safety commitments are insufficient because competitive pressure creates a race to the bottom. Their solution: make transparency obligations legally mandatory at the state level. From the about page: "We believe that the AI ecosystem will be stronger, more secure, and more robust if large AI developers are legally required to [publish safety and security protocols], if whistleblowers who reveal unsafe or noncompliant practices are protected from retaliation, and if developers have clear incentives to mitigate risk."

The intellectual foundation is Thomas Woodside's February 2024 CSET article arguing Congress should codify Biden-era AI reporting requirements into law, since "a future president could decide to rescind the entire executive order." When federal legislation proved impossible and Trump did rescind the Biden order, SAIP pivoted to state-level implementation.

The strategic concession: after SB 1047 was vetoed for including liability provisions, kill switches, and deployment prohibitions, SAIP supported the stripped-down SB 53 -- a transparency-only framework. The logic is that disclosure creates legal records usable in future litigation, establishes industry standards that judges can reference, and provides a floor on which stronger regulation can later be built. But the concession was driven by political reality, not by a belief that disclosure is sufficient.

What They Do

SAIP is a legislative advocacy operation -- not a think tank, not a research org. In 15 months of existence, it has supported bills in 9 states:

California SB 53 (signed Sep 2025): Requires frontier AI developers ($500M+ revenue, 10^26 FLOPs threshold) to publish safety frameworks, release transparency reports before model deployment, report critical safety incidents within 15 days, and maintain whistleblower protections. Penalties up to $1M per violation. Includes CalCompute public cloud cluster. SAIP co-sponsored alongside Encode AI and Economic Security Action California.
New York RAISE Act (signed Dec 2025, significantly weakened): Originally the most ambitious state AI law. Gov. Hochul completely rewrote it under industry pressure. Penalties reduced from $10M/$30M to $1M/$3M. Reporting window extended. "Unreasonable risk" deployment prohibition removed. Creates $9M AI regulatory office.
Utah HB 286 (killed Mar 2026): SAIP hired former Republican House Speaker Greg Curtis as lobbyist, co-commissioned polling with Encode AI showing 91% bipartisan support. Bill passed committee unanimously. White House called it "unfixable" and opposed it publicly. Died without floor vote when session ended.
Six other states (Nebraska, Tennessee, Michigan, Illinois, Massachusetts, Colorado): Bills in committee or withdrawn. Limited reporting on SAIP's specific role.

Sponsors across all bills: 5 Democrats, 4 Republicans. SAIP consistently co-operates with Encode AI across all states.

SAIP produces zero publications, blog posts, research, or public commentary. Their only public statements are quotes in co-sponsors' press releases. They also recruit for government implementation roles -- an anonymous LW forum post directs candidates to implementation@secureaiproject.org for CA and NY enforcement positions.

Key People

Nick Beckstead (CEO, co-founder): PhD Philosophy Rutgers; Open Philanthropy early employee 2014-2021; CEO of FTX Future Fund Nov 2021-Nov 2022 (resigned during FTX collapse); CAIS Policy Lead; FHI fellow; co-founded Giving What We Can's first US chapter. A TIME investigation found he was personally warned about SBF's dishonesty by multiple people in 2018 before going on to lead SBF's philanthropic arm in 2021. His official SAIP bio lists "consultant and grantmaker" for the FTX period without mentioning FTX by name.

Thomas Woodside (co-founder, Senior Policy Advisor): BS Computer Science Yale; Georgetown CSET Horizon Junior Fellow; CAIS policy analyst; worked on SB 1047 at CAIS Action Fund. Co-author with Dan Hendrycks on "Overview of Catastrophic AI Risks." His CSET work on federal reporting requirements is the direct intellectual predecessor to SAIP's legislative approach.

Howie Lempel (Director, Policy): JD Yale Law; former CEO of 80,000 Hours; Interim CEO of Effective Ventures Foundation UK during FTX crisis; former Open Phil Senior Advisor.

Team of approximately 8 people, all with deep EA institutional backgrounds (Open Phil, CAIS, 80K Hours, EVF, CSET, FHI). Several have national security backgrounds (former Army officer, former Naval officer).

Money and Incentives

Financial opacity is total. No 990 filings exist (org too new, EIN ruling date 2026-01-01). No Coefficient Giving/Open Phil grants found. No SFF grants found. No donors are named. SAIP states funding comes from "individual donors and nonprofit institutions" and explicitly rejects corporate funding and foreign government funds.

Estimated operating costs: A team of ~8 in San Francisco, paid lobbyists in at least one state, commissioned polling, and multi-state legislative operations suggest annual expenses of $1.5-3M minimum. The source of this funding is unknown.

Personnel connection to Moskovitz/EA ecosystem: Beckstead spent 7 years at Open Philanthropy (Moskovitz's primary vehicle); Lempel ~1 year at Open Phil. The Washington Examiner and The Dossier frame this as evidence of indirect Moskovitz funding. The personnel history is factual; the funding inference is unproven but plausible.

Anthropic alignment: SAIP does not accept corporate funding, but SB 53 codifies practices Anthropic already follows, giving Anthropic a competitive advantage over less-compliant labs (particularly xAI, which released Grok 4 with no safety documentation). This is not corruption but is a structural incentive alignment worth noting.

No corporate funding claim is unfalsifiable: Without financial disclosure, there is no way to verify SAIP's funding claim. The first 990 filing will be the most informative document about this organization.

What Others Say

Stanford Law (academic critique): SB 53's disclosure approach is fundamentally insufficient. "It would not have prevented Chernobyl, but it would have created an appearance of regulatory oversight." Scores 1.75/5 on key governance principles. "Financial disclosure requirements did not prevent the 2008 crash, environmental impact statements have not halted ecological degradation."

Trump administration (federal opposition): December 2025 executive order directs DOJ to establish an "AI Litigation Task Force" to challenge state AI laws. Called Utah's HB 286 "unfixable." This is an existential threat to SAIP's entire state-level strategy.

Washington Examiner / The Dossier: Frame SAIP as part of EA "dark money" operation using child safety as cover for AI regulation in red states. Factual details about Utah lobbying operations are accurate.

Chamber of Progress (industry critique): SB 53 "penalizes the absence of paperwork, not tangible harm." Revenue-based thresholds are arbitrary. Whistleblower provisions could function as a "layoff shield."

Technical legal critique (Substack): SB 53 confuses risk with impact -- no probability threshold for "catastrophic risk." Risk = probability x impact, but the law omits the probability term, creating enforcement ambiguity.

American Prospect: Hochul "completely rewrote" the RAISE Act with Big Tech-friendly language under pressure from VC Ron Conway and Tech:NYC. Leading the Future super PAC (a16z, Brockman) spent to kill the bill. Shows structural vulnerability of SAIP's approach to industry lobbying.

Dan Hendrycks, CAIS (supportive): "We need much more rigorous regulation to manage AI risks, SB 53 -- and Anthropic's public support for it -- are an encouraging development."

What's Absent

All financial information: No donors, no amounts, no 990 filings, no board members disclosed. An organization whose mission is AI transparency is completely opaque about its own governance and funding.
No candid Beckstead interview about SAIP: He has never publicly discussed why he founded SAIP, his theory of change, or how the FTX experience shaped his thinking. His SAIP bio omits FTX entirely.
No publications or intellectual output: Despite employing some of the most policy-experienced people in the AI safety ecosystem, SAIP produces zero public analysis.
No social media presence: No Twitter/X, no blog, no Substack.
No response to substantive criticism: SAIP has never publicly engaged with arguments that disclosure-only regulation is insufficient.
No CAIS relationship explained: Both co-founders came from CAIS, but the current organizational relationship is undisclosed.