AI models can be dangerous before public deployment (METR, November 13, 2024)

Summary

METR argues that AI models pose significant safety risks during the pre-deployment phase, not just after public release. The post highlights that dangerous capabilities can be elicited from models before they reach the public, making pre-deployment evaluation and safeguards critical. This challenges assumptions that safety interventions only need to focus on publicly deployed systems.

Key Points

• •Dangerous capabilities in AI models can be exploited before public deployment, not just after release to end users.
• •Pre-deployment access (e.g., by researchers, developers, or via API) creates real threat vectors that safety frameworks must address.
• •Current safety thinking may underweight risks from models that haven't yet been publicly released but are already accessible to some actors.
• •METR advocates for robust evaluation and risk assessment processes applied throughout development, not only at the point of public launch.
• •The post emphasizes the need for safety protocols and governance structures that account for the full model lifecycle.