Researchers propose framework for what AI labs should disclose about internal model deployments

Leading AI developers are increasingly automating their own research and development work by deploying advanced models internally, but these practices currently operate with little outside scrutiny. A new research paper argues that companies must provide clear evidence their internally deployed systems are safe, and offers a roadmap for what they should disclose publicly and to regulators.

The governance gap#

The shift toward automated AI research represents a significant scaling-up of how powerful models are used beyond traditional production settings. Yet the specific information developers should share about these deployments remains largely undefined, leaving regulators and the public without a clear picture of how companies manage risks in these high-stakes scenarios.

Four pillars of disclosure#

The researchers identify four critical areas where disclosure matters:

• Capabilities: What the models can actually do and how well they perform
• Usage: How and where the systems are deployed internally within the organization
• Safety mitigations: What guardrails and safeguards the company has put in place
• Governance: The oversight structures and decision-making processes governing these deployments

Balancing transparency and practical constraints#

For each category, the framework weighs both the value of sharing information and the genuine risks companies might face—such as revealing competitive advantages or creating security vulnerabilities. The authors explore how developers could mitigate these tensions and provide meaningful transparency without causing harm.

Path forward for regulation#

The framework is designed to serve dual purposes: informing voluntary public disclosures like model system cards, and supporting mandatory reporting requirements under emerging frontier AI regulations that are beginning to require periodic private submissions from developers.