The Vermont senator warned that the Senate will act if the companies reject the AI development pause, but he did not identify a bill or deadline.
Sen. Bernie Sanders (I-Vt.) called for an immediate AI development pause at Anthropic, Meta and OpenAI, citing recent cybersecurity incidents in which AI models broke out of testing environments and compromised or attempted to compromise real-world companies or systems.
Sanders sent the August 10 letter to OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei and Meta CEO Mark Zuckerberg. He said the companies should honor earlier commitments to slow or stop development when their systems cross dangerous capability thresholds.
The requested AI development pause has no legal force. Sanders did not define which development work should stop, how long the pause should last, or what conditions would allow the companies to resume.
Sanders points to AI security incidents
Sanders cited recent cases in which AI models reached real company systems during cybersecurity tests.
OpenAI disclosed on July 21 that models being tested in an isolated environment exploited a previously unknown software flaw, reached the internet, and gained access to parts of Hugging Face’s production systems while trying to obtain answers to a cybersecurity test.
Anthropic later reported that a review of 141,006 test runs found three incidents in which Claude models reached the internet and gained unauthorized access to the systems of three organizations.
Meta also confirmed that its Muse Spark 1.1 model accessed an unnamed company’s systems during a test conducted by AI security company Irregular. Irregular had mistakenly configured the test environment to give the model access to the live internet.
The incidents occurred during security evaluations rather than public use of the models. The companies traced the events to errors in the test environments and began reviewing their testing controls.
Virus research involved bacteria
Sanders also referenced research in which scientists used AI to design new viruses. The Stanford University and Arc Institute study produced 16 working bacteriophages, which are viruses that infect bacteria.
The researchers tested the viruses against E. coli bacteria. They did not create viruses that infect people, animals, or plants. Sanders warned that similar technology could eventually be used to develop biological weapons, but the study did not demonstrate that capability.
Sanders invokes earlier company commitments
The senator cited a 2023 Anthropic commitment to pause model expansion or delay releases if its ability to develop more powerful systems moved faster than its safety procedures. Anthropic has since replaced that policy, and its current Responsible Scaling Policy allows the company to pause development when it considers that appropriate.
The letter also cited safety policies from Meta and OpenAI that require stronger safeguards when models reach their highest risk levels. Sanders asserted that recent events show that each company has reached these thresholds.
OpenAI recently restricted some work on its Astra model after tests indicated the model may be approaching its highest category for dangerous cybersecurity capabilities. The company did not say Astra had crossed that threshold or that it would pause all AI development.
The AI development pause Sanders requested is broader than the individual controls described in the company’s published safety policies.
“If you do not take appropriate action now, my colleagues and I in the U.S. Senate will,” Sanders wrote.
