Responding to the Risks of Open-Weight AI Models

RUSI has published a resource titled "Responding to the Risks of Open-Weight AI Models," addressing the risks associated with open-weight artificial intelligence models. The resource notes that standalone safety classifiers exist. It states that nothing, however, compels an open-weight provider to run them, meaning a classifier's availability does not establish that it is in use. The resource also observes that many providers supply access to the same models.

Because of that distribution, the resource writes, a determined malicious actor can search across providers for a deployment without safeguards, rather than being limited to any single one. RUSI cites the UK's National Cyber Security Centre, which assessed that AI will "almost certainly" continue to make cyber intrusion operations more effective. The centre also assessed that the proliferation of open source and commercially available models will lower the barrier to entry for a widening range of actors.

The resource places the two observations side by side: safety classifiers are available but unmandated, and the same models are reachable through multiple providers, at least one of which may operate without safeguards. The assessments it cites from the National Cyber Security Centre address both the effectiveness of AI in cyber intrusion operations and the widening range of actors able to enter that space as open source and commercially available models proliferate. Together these points form the case RUSI sets out in the resource on the risks of open-weight AI models.