The Race — Building Democracy in the Age of AI. People race toward a shared finish line.
The Race
← September 2026 overview

SEPTEMBER 2026 · LANE 02

National Security and Frontier Control

Who can access the most powerful AI, and who oversees its use?

Direction of power
Leaning concentration
Window
Narrows
Confidence
Medium-high
Movement since August
Provider-controlled cyber access and state research restrictions deepen; defensive capacity and dialogue partially counterbalance

End-of-September assessment

Advanced frontier capability remains governed through selective access rather than a general entitlement. OpenAI's Astra designation and Google's Fairwind regime make qualification, monitoring and intervention consequential operating decisions. Both distribute defensive capability to approved organizations, but they leave access criteria and expansion under provider control. These actions reinforce the July–August movement toward trusted recipients, controlled releases and specialized state–company arrangements.

Taiwan's cross-government cybersecurity task force builds public capacity to use and oversee frontier systems. The U.S.–China dialogue adds a formal communication route, though an operating incident-notification system is not verified. Neither counterweight changes the basic allocation of advanced capabilities. The September 30 MI5 warning brings existing national-security authority to bear on specified academic relationships. It is not a new ban on AI research or all Chinese partnerships, and intelligence allegations remain attributed.

Public evaluation remains a consequential counterweight. CAISI published an open-weight GLM-5.3 cyber assessment September 17. Its public comparisons demonstrate evaluation output. They do not establish unrestricted pre-release access, adequate staff or assured independence; the quarter-end review in the September synthesis keeps those questions separate.

Other evidence. The Bank's financing-risk discussion is supporting context, not a new rule. The Anthropic–Accenture embedded-evaluation entry is retained at its approved score, but a readable underlying agreement and its independence terms remain an evidence gap. Frontier testing institutions themselves can concentrate specialized access while giving the public more knowledge.

What would change this assessment. Published, reviewable access criteria; independent appeals; public-interest access beyond selected partners; durable external evaluation rights; documented continuity and independence of CAISI; actual use of the interstate notification mechanism; or materially broader legal restrictions on research relationships.

Further September access evidence. The reported request to give U.S. reviewers access before British testers remains unconfirmed as an operating restriction. Anthropic's September 25 D.C. Circuit supply-chain ruling remains a conditional scoring candidate: the March-event comparison is missing, so no additional score is entered. Proposed NIST testing legislation was blocked in unanimous-consent consideration on September 29; the proposed institutions were not enacted. Reported Tencent–Oracle access is attributed and independently unconfirmed, with execution date unknown. None of these holds is cleared by the monthly assessment. See the unscored evidence register for dates, retained dispositions and source links.

Supporting scored events

The complete relevant event list follows. Primary and secondary references preserve one underlying event and score. Source holds and later commencement dates remain explicit; publication does not resolve them.

September scored actionSeptember 30, 2026 evidence cutoff

OpenAI designates Astra Critical in cybersecurity and activates differentiated frontier access

Action/date record: September 1, 2026. Approved score: Concentration +1. Primary lane: 2. Secondary lanes: 5.

OpenAI formally determined that Astra meets the Critical cybersecurity threshold under its Preparedness Framework, the first OpenAI model to receive that designation. Astra demonstrated autonomous vulnerability discovery and exploit development, including a two-zero-day exploit chain. OpenAI plans broad release with more restrictive default safeguards while reserving advanced cyber capability initially for a small tester group and later expanding access through Daybreak Blue.

Sources: Source 1 ↗.

Open the existing event record →

September scored actionSeptember 30, 2026 evidence cutoff

Google launches Fairwind trusted-access regime for advanced cyber capability

Action/date record: September 2, 2026. Approved score: Concentration +1. Primary lane: 2. Secondary lanes: 5.

Google launched the Fairwind Program, providing approved governments, critical-infrastructure operators, software maintainers, cybersecurity companies and other trusted partners access to Gemini 3.8 Flash Cyber, a more permissively safeguarded cyber model than ordinary public Gemini access. Google controls eligibility, permitted uses and the access layer.

Sources: Source 1 ↗.

September scored actionSeptember 30, 2026 evidence cutoff

Taiwan establishes cross-government frontier-AI cybersecurity task force and operating policy

Action/date record: September 4, 2026. Approved score: Dispersion +1. Primary lane: 2. Secondary lanes: 6, 1.

Taiwan’s Administration for Cyber Security, Ministry of Digital Affairs, formally established a cross-government Task Force on Frontier AI Cybersecurity Risks and adopted a phased policy covering frontier-AI cyber-risk monitoring, shared defensive capability, secure-by-design and supply-chain procurement requirements, threat-intelligence coordination, and longer-term domestic AI cybersecurity capacity across government, industry, academia, and research institutions.

Sources: Source 1 ↗.

Open the existing event record →

September scored actionSeptember 30, 2026 evidence cutoff

OpenAI launches Daybreak for Frontline Defenders

Action/date record: September 3, 2026. Approved score: Dispersion +1. Primary lane: 6. Secondary lanes: 7, 2.

OpenAI committed $1 billion in subsidized Daybreak cybersecurity access, training, technical assistance and partnerships for state and local governments, utilities, health organizations, community banks, nonprofits and other essential-service defenders, including a U.S. pilot with MS-ISAC.

Sources: Source 1 ↗.

September scored actionSeptember 30, 2026 evidence cutoff

Australia opens consultation on mandatory national AI-data-center standards

Action/date record: September 2026; processed September 18. Approved score: Dispersion +1. Primary lane: 1. Secondary lanes: 3, 2.

Australia opened a formal standards process covering nationally significant AI and data-center infrastructure. The consultation creates an upstream institutional channel now; the final standards must be assessed separately for binding obligations, enforcement and concentration effects.

Limits and interpretation: Later verification of the operating consultation page puts submission opening at September 18, 06:00 AEST, and closing at October 9, 17:00 AEDT. An indexed account gives September 17. Proposed obligations are not adopted standards. The October 3 Watchlist update adds lanes 6 and 9 without a second score.

Sources: Source 1 ↗; Source 2 ↗.

September scored actionSeptember 30, 2026 evidence cutoff

Anthropic and Accenture establish embedded external model evaluation

Action/date record: September 2026; processed September 18. Approved score: Dispersion +1. Primary lane: 1. Secondary lanes: 2.

The agreement places an outside evaluator inside a continuing frontier-model development relationship rather than relying only on episodic voluntary testing. The score is limited because evaluator appointment, access, publication rights, independence and the treatment of adverse findings remain partly controlled by the provider relationship.

Source hold: The approved ledger refers to an alert source archive, but an operative URL was not recovered from the records materialized for this review. The score is preserved; actor, precise action date, text and scope need source-archive reconciliation before this becomes fully verified public evidence.

September scored actionSeptember 30, 2026 evidence cutoff

California combines enacted AI safeguards with mandatory executive implementation

Action/date record: September 2026; processed September 18. Approved score: Dispersion +2. Primary lane: 1. Secondary lanes: 2.

California combined SB 813 and AB 1405 with Executive Order N-9-26, creating an enforceable package rather than a statement of intent. The package establishes statutory and executive implementation duties, reporting milestones and institutional responsibilities that reduce exclusive provider control over consequential AI-risk information.

Source hold: The approved ledger refers to an alert source archive, but an operative URL was not recovered from the records materialized for this review. The score is preserved; actor, precise action date, text and scope need source-archive reconciliation before this becomes fully verified public evidence.

September scored actionSeptember 30, 2026 evidence cutoff

United States and China establish a formal AI-safety dialogue and proposed incident-notification channel

Action/date record: September 21, 2026. Approved score: Dispersion +1. Primary lane: 2. Secondary lanes: 1.

Senior U.S. and Chinese officials opened a formal AI-safety dialogue and advanced a notification mechanism for AI incidents with national-security implications.

Limits and interpretation: A dedicated dialogue is recorded; an operating incident-notification channel is not established by the available evidence.

Source hold: The approved ledger refers to an alert source archive, but an operative URL was not recovered from the records materialized for this review. The score is preserved; actor, precise action date, text and scope need source-archive reconciliation before this becomes fully verified public evidence.

September scored actionSeptember 30, 2026 evidence cutoff

New York operationalizes the RAISE Act frontier-model oversight framework

Action/date record: September 22, 2026. Approved score: Dispersion +1. Primary lane: 1. Secondary lanes: 2.

New York moved the previously enacted RAISE Act toward operation through developer registration, published safety-framework and 72-hour critical-incident reporting requirements, backed by a Department of Financial Services oversight office and Attorney General enforcement.

Limits and interpretation: The retained Governor source concerns enactment. It does not alone verify the separately claimed September 22 operational milestone; the implementation-source comparison remains open.

Sources: Source 1 ↗.

September scored actionSeptember 30, 2026 evidence cutoff

D30-03 / RSA-026 — UK uses national-security authority to constrain specified academic research relationships

Action/date record: September 30, 2026. Approved score: Concentration +1. Primary lane: 2. Secondary lanes: 1. Stable record identity: D30-03 / RSA-026.

MI5 issued a public espionage alert identifying the China General Technology Research Institute (also translated as the China Academy of General Technology) as closely linked to China's Ministry of State Security. MI5 alleged that the organization funded research by more than 100 UK-linked academics in AI, cybersecurity, covert communications and steganography to assist espionage. These are intelligence-service allegations, not findings independently established by The Race; the reviewed reporting notes China's rejection of espionage accusations.

The alert advised universities to review ongoing and planned relationships and warned of potential criminal consequences under the National Security Act 2023. Security Minister Dan Jarvis said he had written to university vice-chancellors asking them to support staff in ending arrangements with this organization. The action applies existing authority to specified relationships; it neither creates a new law nor prohibits AI research or all Chinese academic partnerships. No prosecution or specific university termination was verified.

Limits and interpretation: The combination of an official intelligence alert, ministerial intervention and potential criminal liability puts state national-security authority behind a constraint on universities' choice of research collaborators. The modest concentration score recognizes this concrete exercise of existing authority without assuming a general research ban or awaiting every eventual enforcement outcome. Concentration describes greater state control; it does not determine whether the security intervention is justified. Being a first encounter with such a warning is not the scoring basis.

Window of contestability: Narrows for the specified academic relationships. The extent and practical effect remain to be observed.

Sources: Source 1 ↗; Source 2 ↗.

Open the existing event record →

Read the additional unscored supporting cases and research →

These findings inform interpretation without receiving event or trend scores.

← All nine lanes