Written by: Li Jia
Anthropic's AI model was halted by the US government within days of its public release, triggered by its main investor, Amazon.
According to the Wall Street Journal, Amazon CEO Andy Garcetti called U.S. Treasury Secretary Bessant and other high-ranking officials, claiming that Amazon researchers had bypassed the security barriers of Anthropic's flagship model, Fable 5, using specific prompts to obtain sensitive information that could be used for cyberattacks. The White House immediately convened an emergency meeting, and the Commerce Department ultimately imposed export controls on Fable and Mythos models, prohibiting access by overseas users . Anthropic subsequently suspended all user access to both models to comply with compliance requirements.
This regulation not only casts a shadow over Anthropic's prospects as it prepares for its IPO, but also reflects the covert power struggle among tech giants, the US government's security anxieties about cutting-edge AI, and the long-standing tensions between Anthropic and the Trump administration.
Amazon's late-night phone call triggers a chain reaction in AI regulation.
According to Axios, the incident began late Thursday night local time. Amazon contacted U.S. government officials and submitted a report stating that its researchers had successfully bypassed the security restrictions of Anthropic's Mythos model, gaining access to content deemed a national security threat.
According to the Wall Street Journal, an Amazon report shows that, guided by specific prompts, the Fable 5 model was able to uncover security vulnerabilities in at least four software programs, information that is usually shielded by the model's security mechanisms.
Andrew Morris, founder of cybersecurity firm GreyNoise Intelligence, points out that this information is still quite far from truly dangerous cybersecurity intelligence, and many publicly available tools can also discover similar vulnerabilities . Furthermore, there is no evidence that Amazon researchers have breached the deeper defenses of the Fable model that transform vulnerability information into executable attack code.
Anthropic responded that the vulnerability pointed out by Amazon is a relatively basic issue, and the company has established adequate security safeguards and always prioritizes security.
Negotiations failed early Friday morning, and Trump signed export control orders.
According to Axios, government officials held several hours of negotiations with Anthropic early Friday morning, demanding that it voluntarily remove its latest model from shelves, but no agreement was reached.
According to The Wall Street Journal, Anthropic CEO Dario Amodei further fueled the government's concerns during a phone call with some officials on Friday. Officials believed that Anthropic's reluctance to cooperate with government security experts to resolve issues corroborated long-held perceptions of its lack of credibility in managing security risks.
Ultimately, despite reservations about restricting innovation, President Trump signed the order approving the restrictions. White House AI advisor and venture capitalist David Sacks stated on social media that the restrictions were "reluctant," adding that the government's "hope now is that Anthropic fixes security issues, export controls are lifted, and the Fable model is re-released publicly."
It is worth noting that Anthropic notified the US government multiple times before the release of the Fable model and collaborated with key government AI testing agencies to adopt a phased release strategy, prioritizing providing preview access to technology companies such as Amazon so that it could be used for software vulnerability screening before being released to a wider audience.
With intertwined interests, Amazon is both a shareholder and a "whistleblower."
The complexity of this incident lies in the fact that Amazon is both a major investor in Anthropic and its chip supplier, while also deploying Anthropic's best models in its own software vulnerability identification business.
According to the Wall Street Journal, citing sources familiar with the matter, Amazon CEO Eric Garcetti's initial call to government officials was a general security warning, but the situation quickly escalated into a comprehensive ban on overseas users by the Department of Commerce . An Amazon spokesperson stated that as a leading cloud service provider serving a large number of public and private customers, it is not uncommon for governments to consult with them regarding potential security risks, and added that the company does not disclose details of such discussions in such cases.
Some analysts point out that the White House's distrust of Anthropic is long-standing , stemming from factors such as the company's ties to liberal political donors, its public warnings about the dangers of AI, and its previous hiring of several Biden administration officials.
Valuation pressure and user diversion could make OpenAI a potential beneficiary.
The timing of these regulations, implemented at a crucial juncture in Anthropic's IPO push, is particularly sensitive.
The company reportedly planned to go public as early as this fall, and the forced removal of its flagship model could lead to users migrating to other platforms and put pressure on the company's valuation. OpenAI is seen as a potential beneficiary , as it has its own robust cybersecurity model, which it is gradually opening to customers, and it is maintaining active communication with the Trump administration.
From a broader perspective, this event, combined with the recent executive order granting security officials greater regulatory authority over AI models and the government's discussions on potential investments in AI companies, signifies a significant escalation in the government's control over the AI industry .
