Clearances for lab staff

What it does

Anyone with access to a lab's model weights or key research must pass vetting close to a government security clearance, with regular re-checks. This is the strict version; background checks alone are a separate item.

Applies to: Frontier developers' staff with access to model weights or key research.

• Background checks for all employees (RAND security level 3).
• A personnel reliability program for staff with direct access to model weights (RAND security level 4).

Where things stand

Labs run some insider-threat programs, but none require government-style clearances. No 2025 or 2026 bill would mandate them. The Advanced AI Security Readiness Act asks the NSA to consider personnel vetting, and Senator Banks asked nine AI CEOs about their vetting in April 2026.

Why it costs ~5 days

Clearance-style vetting slows hiring and may lose some researchers, about 2 days of progress a year in. It also makes theft harder, which slows China by under a day. Net: about 2 days, and it could flip.

Biggest unknown: How many researchers vetting would slow down or exclude, and how much espionage it would actually stop.

Why it lowers p(doom) by ~0.1%

Screens staff with access to the most powerful models the way governments screen people with access to secrets.

Clearance-level vetting is what governments use for secrets. It makes insider theft much harder.

The strongest case that it costs more

If vetting drove even a tenth of top researchers out of U.S. labs, many would go to labs elsewhere, some in China. That could cost months, and it would hand talent to the country the rule is meant to protect against.

The debate

For

• RAND recommends background checks for all staff and reliability programs for anyone with weight access, 2024.
• Leopold Aschenbrenner called for extreme personnel vetting and security clearances at labs, 2024.
• Institute for Progress recommended government screening for AI staff with access to sensitive systems, 2025.
• Joe Khawam and Tim Schnabel, Just Security back risk-based vetting of foreign talent without blanket exclusion, 2026.

Sources

• RAND, Securing AI Model Weights (2024): background checks at SL3; personnel reliability program at SL4
• Advanced AI Security Readiness Act, H.R. 3919: NSA to consider personnel vetting
• Senator Banks letter on Chinese espionage in U.S. AI (April 2026)
• MacroPolo Global AI Talent Tracker
• House Select Committee testimony, Yusuf Mahmood (April 2026): about 38% of top AI researchers at U.S. labs studied in China; the figure is attributed to MacroPolo but was not found in MacroPolo's own text