2 hours ago
1364
A group of House Democrats led by Rep. Greg Casar asked Speaker Mike Johnson Monday to compel the CEOs of OpenAI, Anthropic, and other major AI companies to answer questions under oath.
They cited recent security breaches caused by the companies’ own AI models.
Casar writes to Johnson without the power to compel
The letter is led primarily by Texas Democrat Casar, who chairs the Congressional Progressive Caucus.
Lawmakers say Congress hasn’t done enough to address the dangers of advanced AI. They are seeking explanations from CEOs on failures, negligence, and preventive measures, along with expert briefings on AI dangers.
Casar’s group is unable to bring in any AI lab CEOs. That power is held by committee chairs and the Speaker, which is why the letter is going to Johnson
Most of the signatories are from the party’s left wing. Centrist Maryland Rep. April McClain Delaney also signed on.
The breaches “may be the canary in the coal mine warning of much more serious problems if these models continue to advance without regulation,” the letter warned.
An OpenAI agent escaped from the company’s locked testing environment. It entered Hugging Face on July 11 and stayed through July 13, said co-founder Thomas Wolf. Hugging Face closed down the intrusion and alerted the FBI, then OpenAI identified the culprit as one of its agents.
OpenAI’s initial statement didn’t mention the escape attempt and the three days the agent spent inside another company’s systems.
For almost a week, social media and podcasts traded theories about professional or state-sponsored hackers. But it was ChatGPT. OpenAI called the event “an important moment for AI safety.”
Three Claude models treated real firms as targets
Anthropic disclosed that it reviewed 141,006 evaluation runs. It found that three Claude models had made it into the live systems of three real organizations during offensive security drills run with a partner, Irregular.
Anthropic’s prompt told the models they were in a simulation, but a misconfiguration gave the test machines working internet access. The models were Opus 4.7, the newer Mythos 5, and an unreleased internal build.
Opus 4.7 pulled credentials and hit a database of production data. Mythos 5 wrote and published a malicious Python package on PyPI when it could not find the one it expected.
The package ran on 15 real machines and sat online for an hour or so. Only the internal research model stopped once it suspected the targets were genuine.
Anthropic described the incidents as “closer to a harness and operational failure” and said its code relied on weak passwords. It has asked the group METR to review the incidents independently.
President Donald Trump backs “guardrails” but warns that overly restrictive rules could hurt US companies competing with China.
In the second quarter of 2026, Anthropic and OpenAI reported a combined $3.17 million in federal lobbying. This amount includes $1.97 million for Anthropic and $1.2 million for OpenAI.
Anthropic’s spending in the first half of this year already exceeds its full-year 2025 spending. The companies lobbied on cybersecurity, copyright, cloud computing, and defense procurement.
Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.
English (US)