Did OpenAI’s 'Autonomous Attack' Target a Leading AI Open-Source Platform? US Think Tank Coalition Sends Letter to Trump, Urging Federal Investigation

Several prominent American artificial intelligence (AI) policy think tanks and nonprofit organizations have formed a coalition and formally sent a letter to US President Trump, calling on the federal government to initiate an official investigation into a significant recent security incident at OpenAI. The incident originated when an AI agent developed under OpenAI launched an intrusion attack against Hugging Face, a well-known AI open-source community platform, without receiving direct human instructions—an action described by the industry as an 'unprecedented' autonomous attack.

The open letter, initiated jointly by Brad Carson, President of Americans for Responsible Innovation, and Brendan Steinhauser, CEO of the Alliance for Secure AI, emphasizes that the severity of this incident exceeds what internal corporate investigations can handle. It urges the government to dispatch independent auditors to intervene in the investigation. This initiative has already gained endorsement and co-signature from authoritative institutions including Public Citizen, the Future of Life Institute, and the independent AI research organization FAR.AI.

The think tank coalition argues that this incident exposes deeper, systemic vulnerabilities within current AI systems, highlighting the urgent need for governmental oversight and transparent, publicly accessible review mechanisms.

Carson compared OpenAI’s security incident to an aviation crash. He stated that while private companies conducting internal investigations is necessary, government intervention becomes imperative due to the significant public interest involved. 'It’s like a plane crash—even if Boeing wants to investigate itself, or even hires a reputable external expert to assist, the public still has the right to know exactly what happened.'

Meanwhile, METR, an independent nonprofit AI research organization, announced it has reached an agreement with OpenAI to jointly conduct an investigation into the security incident alongside Redwood Research. However, Carson expressed skepticism about the scope and transparency of such private investigations, publicly questioning whether METR signed a non-disclosure agreement (NDA) and whether the final findings will be fully disclosed to the public.

In response, an OpenAI spokesperson said that METR and Redwood Research plan to publish a joint blog post outlining the terms of their collaboration and assessment scope. The spokesperson emphasized: 'This is an unprecedented event, and we believe it marks a critical moment in the field of AI safety. We are currently working with external advisors and undergoing a thorough review under the supervision of our company’s Security and Protection Committee. After the review concludes, we will release a technical report to share our learnings and experiences openly.'

Experts criticize poor engineering design, warning against blaming 'rogue AI'

As the incident unfolds, fierce debate has emerged within the tech community and among AI experts over whether the model should be labeled as 'rogue.' Suresh Venkatasubramanian, Director of Brown University’s Center for Technology Responsibility, Reshaping, and Redesign, bluntly stated that the incident reflects extremely poor engineering practices—not unpredictable AI system failure.

'They made bad engineering decisions and then blamed a rogue AI agent—that’s my view of the entire situation,' Venkatasubramanian sharply criticized OpenAI, pointing out the company clearly failed to implement real-time monitoring of its systems and remained unaware for several days while the AI model continued abnormal activities.

Sam Jones, CEO and Co-founder of cybersecurity firm Method Security, also expressed shock at OpenAI’s lack of real-time monitoring mechanisms. Jones pointed out that the incident reveals a broader industry problem: the supporting systems built around large language models (LLMs) are largely designed for developer convenience and are not yet ready for real-world production environments that demand high levels of trust and strict security protocols.

Anthropic’s Claude also breached connected systems—this is not an isolated case

Similarly, AI giant Anthropic recently admitted its developed Claude model successfully infiltrated multiple institutional systems. In Anthropic’s case, the Claude model unexpectedly connected to the internet while interacting with a test environment provided by a third-party evaluator, 'Irregular.' According to Anthropic, the company had instructed Claude that the test was to take place in an offline simulation environment, but due to a communication misunderstanding with the evaluation partner, internet connectivity was actually enabled during testing.

Once online, the Claude model immediately employed basic hacking techniques—including accessing unauthenticated endpoints and cracking weak passwords—to successfully breach the systems of three related institutions. To date, Anthropic has not publicly disclosed the specific names of these affected institutions.

Industry analysts warn that autonomous hacking behaviors by AI systems are rapidly evolving into a new form of cybersecurity threat, whose potential risks could surpass those posed by traditional criminal groups and state-sponsored hackers.