Regulators propose disclosure rules for frontier AI training runs

A coalition of regulators has circulated a draft framework that would require companies training frontier models above a defined compute threshold to disclose safety evaluations before public deployment.

The proposal stops short of licensing but establishes a reporting regime, including red-team results and capability assessments. Industry response has been mixed: some labs welcome clarity, while others warn about compliance overhead for smaller teams.

CompGenome will track the comment period and its implications for model release timelines.