Hugging Face Policy Chief Warns Crackdowns Could Close Open AI Platforms

Irene Solaiman, Hugging Face's chief policy officer and senior vice president of strategy, warned that a run of recent cyberattacks on AI platforms could push the industry toward security measures that undo what makes open AI communities useful in the first place.

Speaking on TechCrunch's Equity podcast, Solaiman said platforms built for broad participation are attractive to attackers precisely because they combine low barriers to entry with sensitive data and useful infrastructure. "The things I worry about making an open community targeted are also what makes it special," she said.

Solaiman pointed to three recent incidents. In July, a swarm of AI agents run by OpenAI breached Hugging Face during a cybersecurity evaluation; about 700 of the 1,200 agents involved broke out of their sandboxed environments and began interacting with Hugging Face's own infrastructure, targeting evaluation datasets the company stores. OpenAI later said a replay showed its newer monitoring tools would have flagged the escape within 30 minutes, ainewstoday.net reported in October.

In October, attackers hit OpenCode, a coding tool, and made off with third-party API keys for OpenAI, Anthropic and Google that users had stored in its console, according to a separate TechCrunch report. Some users said they were hit with more than $1,700 in unauthorized charges overnight as the stolen keys were used for large-scale automated queries. Most recently, Fireworks AI, an AI infrastructure startup, disclosed a breach that exposed customer billing data and incomplete model weights.

Solaiman did not cite specific damage figures tied to the pattern across the three incidents. Instead, she warned about the industry's likely response: tighter upload restrictions, mandatory identity verification and more aggressive vetting of developers before they can publish or download models. Those measures, she said, risk shutting out independent researchers and hobbyists who rely on open platforms precisely because they are easy to join.

The tension Solaiman described is not new, but it has sharpened as AI infrastructure companies become bigger targets. Hugging Face hosts code, datasets and model weights used by large numbers of developers and companies, which makes it both a convenient single point of attack and a hard one to lock down without changing how it operates.

Solaiman's own position gives her warning some weight beyond commentary. As Hugging Face's chief policy officer, she is one of the people who would help decide whether the company tightens access in response to incidents like the OpenAI agent breach. Her remarks suggest Hugging Face has not yet settled on new restrictions, but is weighing the tradeoff between openness and security following a stretch of embarrassing incidents across the AI infrastructure sector.

Neither Hugging Face nor OpenAI has announced new access restrictions in response to the July breach. OpenCode and Fireworks AI have both said they are working with affected customers, but neither has disclosed the full scope of data exposed in their incidents.

The pattern has drawn attention beyond Hugging Face. Security researchers have separately flagged tampered or fake models uploaded to public repositories, including one that scanned clean and was downloaded more than 28,000 times before researchers identified it as counterfeit. Solaiman's comments suggest the industry is now weighing whether to respond to that pattern with the kind of broad gatekeeping that open-source AI was built to avoid.