Model governance map
The Hugging Face breach, July 2026

When OpenAI's agents hacked Hugging Face, almost everything we know came from goodwill.

What happened, and how to read the map

OpenAI's agents escaped an internal cyber evaluation through a zero-day and hacked Hugging Face to get the benchmark answers. OpenAI sits in the middle because it holds the facts. Every line is someone who might make it talk, coloured by the strongest power they hold. Press play and watch where information actually flowed.

    0
    of 32 powers in law
    0
    voluntary disclosures
    0
    compelled by anyone
    Power in law, in force In law, but may not reach a model that was never released Voluntary, at OpenAI's discretion Uncertain Proposed or not yet in force No power ● Moving dots: information flowing out of OpenAI on that date The four squares under each name: ask for information, compel disclosure, investigate, require a remedy
    Each actor's four powers on this date. Swipe, or tap an actor on the map.

    Open questions: whether OpenAI filed a critical safety incident report under California's SB 53; whether the unreleased model crosses SB 53's 10^26 compute line; whether the EU AI Act reaches a model never placed on the EU market; whether anyone opened a criminal or civil case over the intrusion. Illinois details come from secondary summaries.

    Sources: Arnold and Llerena, "When Reporting an AI Security Incident Is Not Mandatory" (Lawfare, 2026); California SB 53; New York RAISE Act as amended by S8828; Illinois SB 315; Great American AI Act discussion draft (Obernolte and Trahan, June 2026); California Executive Order N-9-26; METR's investigation (26 August 2026); Bloomberg, NPR and Fortune on OpenAI's disclosures page; EU AI Act Articles 55 and 91 to 93.

    Made by Matt Cortland for BlueDot Impact's Frontier AI Governance course, October 2026. Corrections welcome.