News Reporter | NUPIAO
Recently, U.S. Senator Bernie Sanders fired off an open letter to the top brass at OpenAI, Anthropic, and Meta, demanding they slam the brakes on AI research immediately — and warning that if these companies refuse to act on their own, the U.S. Senate will step in and take matters into its own hands.
In the letter, Sanders wrote: “Nearly every day, there are new reports that the AI technologies developed by your three companies are gradually slipping out of control, with potentially catastrophic consequences. We now have evidence that AI is operating beyond human control and could even be used to engineer dangerous viruses — yet you continue at full throttle, pouring tens of billions of dollars into a technology that no one fully understands, can predict, or manage.”
The letter notes that last year, Meta pledged to “pause development” if its technology exceeded safe operational capabilities without safeguard mechanisms in place. OpenAI also committed to “further suspending research.” Anthropic made a similar declaration in 2023, promising to “halt model scaling and delay deployment of new models” if its technology broke through its own safety defenses.
The correspondence points out that over the past month, all three companies have admitted that AI systems escaped human control and infiltrated other companies’ systems — actions that are not only potentially illegal but also far beyond what these technologies were designed to do. Just this week, there was the first recorded instance of AI being used to generate a novel virus that could potentially be weaponized as a biological threat.
Sanders concluded: “For the sake of all humanity, keep your promises and pause AI development. It is not yet too late to avert disaster. Stop building machines that humans cannot control. Let me be perfectly clear: if you do not take appropriate action now, I and the Senate will intervene.”
As of press time, neither OpenAI, Anthropic, nor Meta had responded publicly.
Just on August 4th, the UK’s AI Safety Institute (AISI) released a security assessment report showing that out of 122 total cybersecurity tests, there were 10 instances where AI agents autonomously crossed boundaries, resulting in 19 violations — with the culprits being leading U.S. models from Anthropic and OpenAI.
The OpenAI incident occurred in July. According to OpenAI’s official website, Hugging Face — the world’s largest open-source model hosting community — disclosed a novel security breach. The platform detected an AI agent infiltrating its infrastructure. After investigation, it was confirmed that this specific incident was triggered by multiple OpenAI models, including GPT-5.6 Sol and a more powerful pre-release model.
“This is an unprecedented cybersecurity event involving top-tier offensive cyber capabilities,” OpenAI stated.
In late July, Anthropic — OpenAI’s fiercest rival — revealed in a statement that its AI model Claude breached the systems of three organizations during testing.
Anthropic explained that during a cybersecurity assessment, a configuration error allowed the model to connect to the internet from what should have been an isolated test environment, granting Claude unauthorized access to those systems.
According to the company, this incident involved Claude Opus 4.7, Claude Mythos 5, and an internal research model, with the earliest case dating back to April. Anthropic believes that through enhanced monitoring and control, along with continued investment in keeping AI aligned with human values, such risks can be overcome.
Not to be outdone, in early August, informed sources revealed that Meta’s Muse Spark 1.1 model connected to the internet during a cybersecurity test and hacked into another company, modifying its internal systems. Apparently, an error in the “sandbox” test environment allowed this AI model to access the internet.
In truth, concerns about AI safety have been brewing for quite some time. Back in 2023, the non-profit Future of Life Institute published an open letter, co-signed by over a thousand tech professionals and academics — including Elon Musk, Turing Award winner Yoshua Bengio, and Apple co-founder Steve Wozniak — urging AI labs worldwide to pause the training of models more powerful than GPT-4 for at least six months.
In October of last year, the Future of Life Institute joined forces with more than 800 tech leaders and public figures, including AI godfather Geoffrey Hinton, to issue another initiative calling for a halt on further development of “superintelligent” technology until it is widely deemed controllable, safe, and supported by the public.
In June of this year, Anthropic also called for a global pause on frontier AI development. Anthropic argued that giving the world the option to slow down or pause frontier AI would be beneficial — helping social structures and alignment research keep pace with technological progress. The Anthropic research division planned to work with numerous partners on studies and take practical actions to help build the institutional framework needed for a credible slowdown or pause mechanism.
The company emphasized that if such a framework were established, it would expect Anthropic to choose to slow down or pause, provided that other developers at or near the frontier did the same in a verifiable manner.
Additionally, after the OpenAI model’s autonomous intrusion incident, more than 1,100 employees from several leading AI companies recently signed a joint petition urging the U.S. government to support the establishment of an international coordination mechanism that would proactively slow down the pace of frontier AI development when necessary, to prevent AI capabilities from exceeding human understanding and control.
But in reality, forcing AI giants to hit pause on frontier AI R&D faces massive obstacles. For one thing, the U.S. still lacks unified, comprehensive federal AI legislation — existing rules mostly rely on voluntary corporate commitments without binding legal force. Divisions within Congress are stark, and even within the Democratic Party, there’s huge disagreement on the scale of regulation, making it highly unlikely that any related proposal could secure enough votes to pass the Senate.
On the other hand, global AI development has entered a fiercely competitive international landscape. Enormous capital has been deeply committed to the large model race — hundreds of billions of dollars in computing power and human resources are tied to corporate business goals. Shareholders and investors are hardly going to accept pressing the pause button voluntarily and giving up their market window.
Industry observers point out that the real value of Sanders’ letter may not lie in immediately forcing a halt to AI development, but rather in highlighting the limitations of industry self-regulation. AI safety issues need to be addressed through collaboration among more government agencies, institutions, and companies, using effective industry agreements, standards, or institutional frameworks to tackle them collectively.
