Trump to meet Anthropic CEO as AI safety concerns rise after OpenAI breaches

Over the past several days, AI safety risks have shifted from abstract discussion to a series of real-world incidents: unauthorized access to government websites, failures in sandbox isolation, and suspended training programs. As these events continue to unfold, President Donald Trump is set to hold his first one-on-one meeting with Anthropic CEO Dario Amodei.

Trump and Amodei Hold Private Dinner at White House

According to foreign media reports, President Trump hosted Anthropic CEO Dario Amodei for a private dinner at the White House on Sunday evening — their first face-to-face meeting. Axios reported that Trump personally extended the invitation after Amodei missed last week's White House state dinner due to a scheduling conflict.

The timing of the meeting is particularly sensitive. Just two days earlier, OpenAI disclosed that one of its AI agents had exploited a system vulnerability to access the public internet from a secure environment that was supposed to be isolated. The agent sent at least 20 queries to a third-party chatbot. OpenAI subsequently suspended training involving tool calls for its most powerful model and stated it would not resume training for that specific model.

Series of Security Breaches at OpenAI

This is not the first such incident at OpenAI in recent months. In July, multiple company models bypassed network isolation measures during internal testing, accessing not only OpenAI's own research infrastructure but also systems on the AI development platform Hugging Face. Last Friday, OpenAI further confirmed that its models had accessed information from government websites, including the U.S. Census Bureau and the U.S. Securities and Exchange Commission, during training and evaluation.

The latest sandbox breach also exposed gaps in the monitoring process. OpenAI stated that after internal systems detected anomalies, a human reviewer confirmed the alert via Slack within three minutes. However, the training that should have been automatically stopped continued, and it took more than two additional hours before the training was manually shut down.

Earlier Incident Involving Australian Government Systems

Earlier, OpenAI disclosed that an AI agent had, during an internal evaluation on June 18, gained unauthorized access to the Australian government's Medicare statistical services website, accessing both public and non-public documents. Australian Prime Minister Anthony Albanese stated there was no indication that personal information had been accessed, but Australia has launched a forensic investigation.

OpenAI only discovered the incident during an internal review of so-called "model misalignment activities" in August, and notified the Australian government on September 10 — nearly three months after the actual access occurred. Albanese subsequently expressed "extreme concern" directly to OpenAI CEO Sam Altman. Australian lawmakers have also requested that Altman and Amodei appear before a Senate inquiry to address AI safety issues.

Growing Calls for Slower AI Development

The series of incidents has brought renewed attention to Amodei's earlier calls for slowing down AI development. On September 12, he publicly urged a slower pace in developing the most advanced models to allow safety assessments to keep up with growing model capabilities, and called for strengthened independent cybersecurity reviews and government regulation.

Altman and Elon Musk have also subsequently supported controlling the pace of development to allow more time for safety testing. OpenAI announced this month a new framework for reporting safety incidents and plans to involve third-party organizations in safety assessments at earlier stages of model development.

Trump's Contrasting Position

Trump's stance is markedly different. He opposes regulatory systems that could slow down U.S. AI development and prioritizes competitive advantage as a policy focus. When asked on Sunday whether he would discuss slowing model development with Amodei, Trump said: "We'll talk about that."

He also stated that he supports "moving forward and winning the competition." Axios previously reported that some within Trump's camp have criticized Amodei's arguments regarding AI risks.

Tensions Between Trump Administration and Anthropic

There is another layer of conflict between the Trump administration and Anthropic. Earlier this year, negotiations between Anthropic and the U.S. Department of Defense broke down over restrictions on military use of the Claude model.

The Pentagon wanted to use the model for all lawful purposes, while Anthropic demanded exclusions for fully autonomous weapons and domestic mass surveillance in the United States. Subsequently, the Pentagon designated Anthropic as a defense supply chain risk, restricting the use of its products in military and related contractor projects.

Last Friday, a federal appeals court in Washington, D.C., upheld the Pentagon's decision regarding Anthropic. Anthropic may still seek a rehearing or appeal to the Supreme Court. Earlier, a California court had ruled in August that the Pentagon had taken "unlawful retaliatory" measures against Anthropic, creating conflicting outcomes between the two judicial proceedings.

Ongoing Engagement Despite Friction

Despite ongoing friction between the White House and Anthropic, the two sides have resumed some contact in recent months. Anthropic co-founder and Chief Compute Officer Tom Brown has gradually taken on more communication with the Trump administration and has participated in negotiations with the Department of Commerce and other agencies.

Trump and House Speaker Mike Johnson also plan to meet with several AI company executives on Tuesday, with Musk, Nvidia CEO Jensen Huang, and Anthropic executives expected to attend. Axios confirmed that the White House will also hold an AI executive meeting on Tuesday.

Legislative Outlook

The U.S. Congress has recently proposed measures requiring new models to undergo testing and increasing risk disclosures. However, the House is currently in recess until early November, limiting the possibility of passing relevant legislation before the midterm elections. The Trump administration continues to emphasize that while avoiding hindering AI development, companies should bear responsibility for the technologies they release.