Center for Long-Term Cybersecurity (CLTC)

Stated Theory of Change

CLTC's AI-safety-relevant theory of change operates through institutional infrastructure. The logic chain:

1. AI systems pose serious risks including CBRN, cyber, model autonomy, and socioeconomic disruption.
2. Risk management must be standardized and made actionable before it can be mandated.
3. Academic institutions can serve as neutral convening platforms to develop these standards with multi-stakeholder input.
4. Standards like the GPAIS Profile create the technical foundation for regulation, compliance, and accountability.
5. Training interdisciplinary researchers creates a pipeline of people who can implement these standards in government, industry, and civil society.
6. Independent academic oversight of risk thresholds prevents companies from being "the only ones deciding what is safe enough."

The AISI states: "We understand that the harms from AI may be devastating and that there is a narrow window of opportunity to meaningfully address them." This is strong language for a mainstream university center -- it acknowledges urgency and catastrophic stakes without using EA/rationalist terminology.

Revealed Theory of Change

CLTC's actions are broadly consistent with its stated mission, but with important nuances:

Standards work is genuine and the output is substantive. The GPAIS Profile took a year to develop with 100+ stakeholders, was retrospectively tested against 4 frontier models, and received endorsements from organizations spanning the x-risk/industry/policy spectrum. The Agentic AI Profile addresses frontier risks (self-replication, resistance to shutdown). The intolerable risk thresholds paper provides specific, measurable limits. This is not box-checking -- it's serious governance work.

The cybersecurity side gets more resources and institutional attention. The Consortium of Cybersecurity Clinics (60+ clinics, $2.2M Google grant, 2,200 students/year) dwarfs the AI Policy Hub (18 fellows total, 3 cohorts). The Cyber Resilience Corps has its own program director and dedicated staff. The AI Security Initiative is run by an interim director (Madkour, who is technically a BERI employee) and is hiring for a permanent Program Director. Resource allocation reveals that cybersecurity is the institutional center of gravity.

BERI serves as the existential risk bridge. The two most AI-safety-relevant people at CLTC (Barrett and Madkour) are both BERI-connected. SFF funding flows through BERI. Open Philanthropy grants to CLTC are paired with companion grants to BERI. This structure lets x-risk-motivated work happen within a mainstream institution without requiring CLTC itself to adopt x-risk framing. It's clever institutional design, but it means CLTC's AI safety work depends on continued BERI involvement.

Talent pipeline is the most clearly demonstrated impact. AI Policy Hub alumni at OpenAI safety (Carroll authored manipulation measurement research, Whitney co-authored safety thresholds for open-weight models), FBI AI policy, Meta security evaluations -- these are traceable, concrete placements. The 86% career influence rating from alumni is strong.

The center is in a leadership vacuum. Four leadership transitions in four years (Weber, Hoofnagle, Newman, Saxenian all departed or stepped back), with the AI-specific leader (Newman) now at Microsoft. This creates genuine risk that institutional knowledge, external relationships, and strategic direction are being lost.

Key Assumptions

Assumption 1: Voluntary standards can evolve into mandatory regulation.

• Evidence for: NIST Cybersecurity Framework started voluntary and became referenced in regulation and executive orders. Sector regulators increasingly reference NIST AI RMF. EU AI Act maps to the GPAIS Profile.
• Evidence against: No US federal AI legislation has passed as of 2026. Karnofsky estimates full risk management maturity could take "decades." Companies can endorse standards without implementing them.
• Testable: Track whether GPAIS Profile or risk thresholds are referenced in binding regulation within 3-5 years.
• If wrong: CLTC's standards work becomes a scholarly exercise with endorsements but no enforcement mechanism.

Assumption 2: Academic independence provides credibility advantage over industry self-regulation.

• Evidence for: "How to Say No" article explicitly argues for independent oversight. CLTC has no financial dependence on any single AI company. University affiliation provides institutional protection.
• Evidence against: Google.org is the largest known single funder ($2.2M). Advisory committee includes reps from AWS, Google, Meta. The GPAIS Profile was endorsed by Google DeepMind -- is that independence or co-option?
• If wrong: CLTC's standards become another form of industry self-validation dressed in academic credibility.

Assumption 3: The cybersecurity-to-AI-safety pipeline works.

• Evidence for: The AISI grew out of cybersecurity expertise. Risk management frameworks are transferable. Cybersecurity community engagement (WEF, Fortinet, NIST) opens doors for AI governance work.
• Evidence against: AI safety requires understanding of alignment, capability elicitation, and emergent behaviors that cybersecurity expertise doesn't provide. CLTC has no technical AI safety research. The "AI Security" framing may not capture catastrophic risk dynamics.
• If wrong: CLTC produces governance frameworks that address the wrong risks or underweight the most dangerous ones.

Assumption 4: CLTC's institutional capacity survives the leadership transition.

• Evidence for: Cleaveland provides organizational continuity. BERI/SFF funding continues. Madkour has standards expertise. Active hiring for permanent AISI director.
• Evidence against: Newman was irreplaceable in terms of external visibility, publication output, and standards body relationships. The Faculty Director position is held by a retired interim. Two consecutive Faculty Directors left.
• If wrong: CLTC's AI safety work stalls or contracts while the cybersecurity programs continue.

Strengths

Unique institutional positioning. CLTC occupies a rare niche: a mainstream academic center with NIST engagement, WEF convening power, and industry credibility, that also produces work endorsed by FLI and GovAI and funded by Open Philanthropy and SFF. Almost no other organization bridges these worlds as effectively. This positioning means CLTC's standards outputs are taken seriously by both the x-risk community and mainstream policy institutions.

Standards that actually address frontier risks. The GPAIS Profile and Agentic AI Profile don't just address bias and privacy. They explicitly cover CBRN, cyber attacks, model autonomy, self-replication, and resistance to shutdown. The intolerable risk thresholds paper provides specific recommendations for these categories. This is frontier AI safety content in standards-body-compatible packaging.

Demonstrated talent pipeline. 18 AI Policy Hub fellows with traceable placements at OpenAI safety, FBI, Meta security, and academic institutions. The "policy accelerator" model turns researchers into practitioners who carry AI safety knowledge into institutions with power.

Iterative approach to standards. The GPAIS Profile has gone through drafts, v1.0, and v1.1 with EU AI Act mapping. They're actively updating it. This iterative approach is exactly what Karnofsky argues for and is more realistic than trying to produce definitive standards.

Part of the Berkeley AI safety cluster. CLTC, CHAI, BERI, BRSL, and the AI Policy Hub form an ecosystem. Stuart Russell advises the AI Policy Hub. Barrett connects GCRI and BERI to CLTC. This cluster effect multiplies each org's influence.

Weaknesses and Risks

Leadership crisis. Four departures in four years, with the most critical (Newman) going to industry. The AISI is run by an interim director who is technically a BERI employee. The Faculty Director is a retired interim. Hiring for a permanent AISI director is ongoing. This is a fragile institutional state.

Financial sustainability is uncertain. The founding $15M Hewlett grant and broader Cyber Initiative ended Dec 2023. Post-Hewlett funding picture is opaque. AI-safety-specific funding (~$800K identified) is small. University centers can die slowly when founding grants expire and no replacement emerges.

Voluntary standards gap. CLTC's own researchers acknowledge that voluntary standards are insufficient. They call for mandatory regulation and enforcement. But CLTC has no power to compel adoption. The theory of change depends on regulators and legislators picking up what CLTC puts down -- and there's no evidence yet of binding adoption.

No adoption metrics. How many companies actually use the GPAIS Profile? How many regulators reference the risk thresholds? CLTC reports endorsements but not usage. Without this data, the theory of change is unverifiable.

Zero EA/rationalist community engagement. CLTC is invisible to the community that thinks most seriously about catastrophic AI risk. This means no feedback loop, no community scrutiny, and potential systematic underweighting of CLTC's contributions in AI safety ecosystem assessments.

No technical AI safety research. CLTC governs risks that others identify. If the technical AI safety community misidentifies the risks, CLTC's standards will address the wrong problems. This dependency is by design (standards bodies don't do basic research) but creates vulnerability.

Industry advisory committee optics. Advisory committee members from AWS, Google, Intel, Meta are explicitly tasked with fundraising. Google is both funder and regulated party. This is common in standards bodies but creates the appearance of capture that could undermine credibility with more skeptical audiences.

Cross-References

Complementary to METR/ARC Evals: METR evaluates dangerous capabilities; CLTC translates those evaluations into standards and thresholds. They address different parts of the same problem.

Complementary to CHAI: Russell's alignment research provides the technical understanding; CLTC's standards work provides the governance infrastructure. Russell advises the AI Policy Hub directly.

Complementary to BERI: BERI provides existential risk framing, funding channels, and personnel (Barrett, Madkour) while CLTC provides institutional credibility and standards expertise.

Adjacent to GovAI: Both work on AI governance from academic positions. GovAI endorsed the GPAIS Profile. Different emphasis: GovAI does more theoretical governance research; CLTC does more applied standards development.

Different approach from PauseAI/advocacy orgs: CLTC works within institutional frameworks (NIST, OECD, ISO) to build governance infrastructure. Advocacy orgs push for political action. These are complementary -- advocacy creates demand for standards, standards create the technical basis for regulation.

What Would Change This Assessment

Upward revisions:

• If GPAIS Profile or risk thresholds are referenced in binding US federal legislation or EU implementing regulation.
• If a permanent AISI director with strong AI safety credentials is hired.
• If major AI company announces specific changes to practices citing CLTC standards.
• If CLTC publishes adoption metrics showing significant industry uptake.
• If post-Hewlett funding is secured at comparable levels from AI-safety-aligned funders.

Downward revisions:

• If AISI director hiring fails or results in a candidate without AI safety expertise.
• If AI-safety-specific funding (Open Phil, SFF/BERI) is not renewed.
• If Newman-caliber researchers don't replace the departed talent.
• If CLTC's standards work is cited by companies primarily as evidence of voluntary compliance rather than as a basis for actual practice changes.
• If the cybersecurity pillar continues to grow while the AI pillar contracts.

Self-Critique

What sources should I have checked but didn't?

• The actual PDF text of the GPAIS Profile (174 pages) and Agentic AI Profile. I read descriptions and summaries but not the full technical documents.
• The Hewlett Foundation summative evaluation of the Cyber Initiative, which would reveal how Hewlett assessed CLTC's performance.
• Newman's LinkedIn for exact departure timing and any public statement.
• UC Berkeley budget documents that might reveal CLTC's allocation.

Where is this analysis potentially biased?

• I may overweight the significance of CLTC's AI safety work relative to its cybersecurity work because the prompt directs me toward AI safety relevance. CLTC's leadership and most staff are primarily cybersecurity-focused.
• The absence of criticism could lead to an overly positive assessment. No criticism found does not mean no criticism exists -- it may mean the org is too small or too outside EA circles to attract scrutiny.
• I infer financial fragility from leadership turnover and the Hewlett exit, but CLTC may have secured funding sources I haven't identified.

What would a thoughtful person who disagrees say? A skeptic might argue: "CLTC is a cybersecurity center that bolted on AI governance to follow the funding. The AI safety work is small, underfunded, dependent on two BERI employees, and just lost its most important person. The standards they produce are endorsed but not adopted. The theory of change requires regulators to act, and there's no evidence they will. The $800K in identified AI safety funding is negligible compared to what technical AI safety orgs receive. Why would anyone focus on CLTC?"

My weakest claim: That CLTC's standards work will evolve into binding regulation. This is the core of the theory of change, and there is no direct evidence of it happening yet. The NIST Cybersecurity Framework precedent is encouraging but not proof.

What information would most change my view? Adoption data. If I could see that X companies have implemented the GPAIS Profile and Y regulators reference the risk thresholds, that would dramatically increase my confidence in the theory of change. Conversely, if 3 years pass with endorsements but no adoption or regulatory reference, I would significantly downgrade CLTC's AI safety relevance.