The public release of Anthropic's AI models was halted by the U.S. government within days, with the trigger coming from its major investor.
According to reports, Amazon CEO Andy Jassy contacted high-level officials including Treasury Secretary Besant, stating that its researchers had bypassed the safety guardrails of Anthropic's flagship Fable 5 model using specific prompts, obtaining sensitive information usable for cyberattacks. The White House subsequently convened an emergency meeting, leading the Commerce Department to impose export controls on the Fable and Mythos models, blocking access for users outside the U.S. Anthropic then suspended all user access to both models to comply with regulatory requirements.
These controls not only cast a shadow over the prospects of Anthropic, which is preparing for an IPO, but also highlight the underlying power struggles among tech giants, the U.S. government's security anxieties regarding advanced AI, and the long-simmering tensions between Anthropic and the current administration.
Amazon's Late-Night Call Ignites Chain Reaction of AI Controls
The incident reportedly began late on a Thursday. Amazon contacted U.S. government officials and submitted a report stating its researchers had successfully breached safety restrictions of Anthropic's Mythos model, accessing content deemed a national security threat.
Further reports indicate Amazon's findings showed that, with specific prompting, the Fable 5 model could uncover security vulnerabilities in at least four software programs—information typically blocked by the model's safety mechanisms.
A cybersecurity expert noted that this information remains some distance from genuinely dangerous cyber threat intelligence, and similar vulnerabilities can be found using many public tools. Moreover, there is no evidence that Amazon's researchers breached deeper protections within the Fable model that prevent turning vulnerability information into executable attack code.
Anthropic responded that the vulnerabilities pointed out by Amazon were relatively basic issues, and the company has established robust safety measures, always prioritizing security.
Failed Negotiations and Presidential Approval of Export Controls
Government officials reportedly engaged in hours of negotiations with Anthropic in the early hours of Friday, requesting the company voluntarily take down its latest models, but no agreement was reached.
During a subsequent call, Anthropic's CEO, Dario Amodei, is said to have further deepened officials' concerns. Officials perceived Anthropic as unwilling to cooperate with government security experts to resolve issues, aligning with the government's longstanding judgment that the company lacks credibility in security risk management.
Ultimately, President Trump, despite reservations about stifling innovation, signed off on the control action. A White House AI advisor described the restriction as a "reluctant" move, expressing hope that Anthropic would fix the security issues, leading to the lifting of controls and the public re-release of the Fable model.
It is noteworthy that Anthropic had notified the U.S. government multiple times before the Fable model's release and collaborated with key government AI testing agencies, employing a phased rollout strategy that prioritized preview access for tech companies like Amazon.com for software vulnerability screening before a wider public release.
Intertwined Interests: Amazon as Both Investor and Whistleblower
The complexity of this event lies in the fact that Amazon.com is a major investor in Anthropic, a chip supplier, and also deploys Anthropic's most capable models within its own software vulnerability identification business.
According to sources, Amazon CEO Jassy's initial intent in contacting government officials was a general security warning, but the situation rapidly escalated into a Commerce Department ban for all non-U.S. users. An Amazon spokesperson stated that as a leading cloud service provider for numerous public and private sector clients, it is not uncommon for the government to consult them on potential security risks, adding that the company does not disclose details of such discussions.
Some analysts point out that the White House's distrust of Anthropic has deep roots, involving the company's connections to liberal political donors, its public warnings about AI dangers, and its previous hiring of several officials from the prior administration.
Valuation Pressure and User Diversion: OpenAI as a Potential Beneficiary
The timing of these controls is particularly sensitive as Anthropic approaches a critical juncture in its push for an IPO.
Reports suggest the company initially planned to go public as early as this fall. The forced removal of its flagship model could lead users to migrate to other platforms and put pressure on the company's valuation. OpenAI is seen as a potential beneficiary, as it possesses its own powerful cybersecurity model, is gradually making it available to customers, and maintains positive communication with the current administration.
From a broader perspective, combined with recent executive orders granting security officials greater regulatory authority over AI models and government discussions about potential equity stakes in AI companies, this event marks a significant escalation in the government's control over the AI industry.
Comments