Why did Representative Ro Khanna ask OpenAI, Anthropic, Google, Meta and SpaceXAI to disclose known attempts by China or other hostile actors to illegally access their AI model weights and describe their cybersecurity safeguards, how does theft of those weights differ from the mo

Khanna wanted to know whether leading U.S. AI companies can keep their most valuable models out of hostile hands—not merely whether rivals can imitate their outputs. His request for reports of attempted illegal access and descriptions of cybersecurity safeguards treats model security as a national-security question that companies should be accountable for answering.510

• The distinction: Stealing model weights means obtaining the model’s underlying learned parameters. Distillation means using a model’s responses to train a different model; it does not, by itself, give the imitator the original weights. OpenAI and Anthropic have reported distillation involving Chinese firms, but the reporting identifies few known cases of stolen weights. Distillation reports therefore are not evidence that those firms’ weights were stolen.10
• The risk: Khanna’s concern is that a hostile actor with the actual weights could gain direct control over a powerful model, beyond access through its developer’s service. His earlier questioning explicitly distinguished distillation from the problem of safeguarding weights.5 The available evidence does not establish that China has successfully stolen the weights of any company named in the question.10
• The policy divide: Trump’s approach has emphasized voluntary industry measures; his administration also said it would not safety-test open-weight models under a voluntary testing arrangement.7 Khanna instead advocates required audits, stronger government oversight and a binding, verifiable U.S.–China AI-safety agreement. His letters press for evidence that company-led safeguards are adequate rather than assuming they are.123

The specific contents of the companies’ safeguards or any undisclosed intrusion attempts are not established by the available evidence.