Checks before lab use
Safety checks before internal use. Test and safeguard a model before the lab's own researchers use it, not just before the public does.
What it does
Before a lab's own staff can use a new frontier model, for example to write code and run experiments, the lab must finish its dangerous-capability tests and put safeguards in place, the same way it would before a public launch.
Applies to: Frontier models, at the point of internal deployment.
- Complete dangerous-capability evaluations before a frontier model is deployed inside the lab.
- Put required safeguards in place before internal use, not only before public release.
- California SB 53 already requires large developers to assess catastrophic risk from internal use and send the state summaries every three months.
Where things stand
Labs already test models before release, and some safeguards now apply internally. OpenAI's GPT-6 Astra system card says its Preparedness Framework required safeguards even for internal deployment because of the model's cyber capabilities. California's SB 53 requires a framework for managing risk from internal use. What labs do not do is hold every new model back from their own staff until testing is done.
Why it costs ~4 days
Researchers keep using the previous model while the new one is tested, so the loss is only the extra help the newest model would have given. About a day over a year.
Biggest unknown: How long labs would hold a model back from their own staff while tests finish.
Why it lowers p(doom) by ~0.25%
Catches dangerous models before labs use them in-house, where they help build the next model.
Many risk scenarios start inside the lab, where a powerful model helps build its successor. Checking models before internal use targets that.
The strongest case that it costs more
As labs automate more of their own research, the internal window becomes the whole race. If a model has to wait for full testing before it can help build its successor, and testing keeps getting longer, the cost could reach a month or more per generation and compound.
The debate
For
- Apollo Research recommended extending frontier safety policies to internal deployment, 2025.
- California SB 53 requires frameworks to manage risk from internal use, 2025.
- Joe Kwon and Stephen Casper argued U.S. and EU rules leave internal deployment under-supervised, 2026.
Sources
- Stix et al., AI Behind Closed Doors: a Primer on the Governance of Internal Deployment (Apollo Research, April 2025): governance of internal deployment of highly advanced systems appears absent
- California SB 53 bill text: internal-use risk in the frontier AI framework; quarterly summaries
- OpenAI GPT-6 Astra system card: safeguards: safeguards required even for internal deployment
- METR Frontier Risk Report, Feb–Mar 2026: internal frontier about 66 days ahead of the public
- Anthropic, When AI builds itself: more than 80% of merged code authored by Claude as of May 2026
- UK AISI, Early lessons from evaluating frontier AI systems
Rough starting points, not precise forecasts. Lead costs assume China doesn't depend on U.S. models, the case least favorable to safety laws, and count 3 years. On the menu you can change every assumption and put in your own numbers. Last priced 2026-09-26.