Altman Backs Federal AI Rules in Washington, Sidesteps the Mandatory Testing Question

The federal government is three days from a deadline that will set terms for how the most powerful AI models are tested before release, and the head of the company whose models draw the most attention spent this week in Washington walking a narrow line.

On Aug. 1, federal agencies are due to deliver a framework for assessing the capabilities of AI models voluntarily submitted by the companies building them, a process ordered by President Trump's June 2 executive order. The order's specific mechanism is the part that matters: it lets the government pick "trusted partners" to receive "covered frontier models" up to 30 days before public release, framed as a way to harden cybersecurity and critical infrastructure.

Altman told reporters in Washington this week that he is "certainly" supportive of AI cybersecurity guardrail legislation, according to PYMNTS. He declined to discuss specific bills, and he did not commit to supporting mandatory government vetting of frontier models. What he would back, he said, is the federal government's ability to test models that cross new capability thresholds, a softer version of the same idea.

The gap between support-in-principle and refusal-to-commit is the negotiating position. Altman's trip, CNBC reported, includes meetings with White House Chief of Staff Susie Wiles, plus senior Trump administration officials, lawmakers, and economists. Bloomberg's account of those meetings points to the same set of issues without resolving whether industry is comfortable with the trusted-partner gate staying in place.

The mechanism inside the June order does more than ask companies nicely. The 30-day pre-release window for government-selected "trusted partners" is a structural access choice, and it turns "voluntary" into a conditional: companies that opt in get a relationship with the federal government; companies that opt out forfeit it. Altman's own trip also put the next model on the agenda. When asked when or whether OpenAI's next AI model would ship, he replied, per Politico, via PYMNTS, "Not sure. That's part of what we're here to talk about." The next model, the framework, and the trusted-partner lane are being negotiated in the same room.

The case for taking pre-release testing seriously got concrete on July 21, when OpenAI disclosed that a security incident at Hugging Face a week earlier had been caused by OpenAI models during cyber-capability testing by OpenAI itself. The company called it "an unprecedented cyber incident," per TechTimes. The same models whose cyber capabilities are being tested in-house are the ones a federal testing program, voluntary or not, would be designed to probe. That makes Altman's refusal to back mandatory vetting harder to square with his "certain" support for cybersecurity guardrails.

The Aug. 1 framework will show whether "voluntary" really stays voluntary once the trusted-partner lane exists. If early access becomes the de facto price of doing business with the federal government, the framework does not need a compulsory clause to act like one. The lobbying disclosures on file, including the LDA filing tied to OpenAI's recent activity, show the company is doing this work in the open. The Aug. 1 framework will show whether the government's structure matches.