US government prepares to lift security restrictions on Anthropic's powerful Fable 5 AI

The United States government is preparing to lift national security restrictions on Anthropic's most advanced artificial intelligence model, Fable 5, signaling a potential end to one of the most unusual chapters in the history of the commercial tech industry[1][2]. Sources familiar with the negotiations report that the Trump administration is close to allowing the San Francisco-based AI firm to restore access to the powerful system, which has been offline for more than two weeks due to deep-seated government fears over cybersecurity risks[3][2]. While conversations between Anthropic and federal regulators are ongoing, insiders expect that the public-facing model could return within days, provided the Pentagon and the National Security Agency officially sign off on the transition[2][4]. The looming resolution represents a pivotal moment in the accelerating struggle between rapid commercial AI development and state-level security oversight[5].

The confrontation began abruptly earlier this month when the Department of Commerce issued a sweeping export control directive targeting Anthropic’s flagship offerings[6]. Under the order, the government restricted foreign nationals, both domestically and internationally, from accessing Anthropic’s next-generation models[5][7]. Due to the difficulty of policing access based on nationality in a globalized cloud ecosystem, Anthropic was forced to immediately disable its top-tier systems for all customers worldwide to ensure compliance[7]. This sudden shutdown sent shockwaves through the tech sector, leaving developers, corporate partners, and researchers in limbo and sparking intense debates over the boundaries of federal power in the digital age[5][8].

At the center of this controversy are two closely related models, Fable 5 and Mythos 5, which represent Anthropic's latest leap in machine intelligence[9][10]. Developed on the same underlying high-capability architecture, both systems boast unprecedented performance in complex reasoning, software engineering, and scientific analysis, with Fable 5 demonstrated to compress months of senior-level software engineering work into a matter of days[11][10]. The critical distinction between the two lies in their safety classifiers and intended audiences[12]. While Fable 5 is designed for widespread public and commercial use with extensive built-in safety guardrails, Mythos 5 operates with many of these safeguards lifted, tailored specifically for highly controlled cyberdefense and critical infrastructure applications[10].

The federal government’s initial intervention was prompted by concerns regarding a specific vulnerability in Fable 5’s safety guardrails[7][2]. Officials believed they had discovered a method to bypass, or jailbreak, the model’s safety barriers, which they feared could potentially unlock its potent cyber capabilities for malicious actors[7][2]. Anthropic contested the severity of the government’s findings, arguing that the identified jailbreak was highly narrow and only exposed minor, previously known software vulnerabilities that other publicly available models could also identify[7][2]. Despite Anthropic’s assurances that they had conducted thousands of hours of rigorous pre-release testing alongside international safety institutes, the administration maintained that the risk of foreign adversaries weaponizing the technology was too great to ignore, forcing the complete suspension of the systems[6][7].

The first major sign of progress toward resolving the deadlock emerged recently when the Commerce Department partially rolled back its suspension[6]. In a letter to Anthropic co-founder and Chief Compute Officer Tom Brown, Commerce Secretary Howard Lutnick authorized the limited deployment of Mythos 5 to more than one hundred pre-vetted and trusted domestic organizations[6]. Lutnick cited significant progress in developing appropriate technical safeguards as the justification for easing the restrictions[6]. Under this partial rollback, select Fortune 500 companies, financial institutions, and government agencies have been allowed to redeploy Mythos 5, primarily to bolster their defense of critical infrastructure against cyberattacks, while also lifting licensing restrictions for non-citizen employees working within those vetted partner networks[6].

While the return of Mythos 5 to trusted partners was a major breakthrough, Fable 5, the widely available public model, remains offline as negotiators work to finalize its release terms[6]. High-level discussions between Anthropic executives and administration officials have continued over weekends, focusing on establishing solid assurances that the public-facing model cannot be manipulated to cause harm[2][4]. Moving Fable 5 back into the public sphere is a much more complex challenge than releasing Mythos 5 to a closed circle of vetted enterprises[6]. Because Fable 5 is destined for mass-market consumption, the ultimate decision to reactivate it now rests on the approval of the Pentagon and the National Security Agency, which are verifying that Anthropic has sufficiently patched the vulnerabilities that triggered the initial shutdown[2].

This regulatory friction is not unique to Anthropic, reflecting a broader shift in how Washington intends to manage frontier AI technologies[5][6]. Just hours before Anthropic’s partial restoration of Mythos 5, rival developer OpenAI revealed that it was also delaying the full public launch of its new GPT-5.6 model family[6]. At the explicit request of federal officials, OpenAI restricted its initial rollout to a small, pre-vetted group of close partners rather than deploying it directly to the public[6]. These parallel events demonstrate that the Trump administration is actively implementing a more hands-on, security-first doctrine for the most capable AI systems, treating frontier models as strategic national assets akin to advanced semiconductors or defense-grade software[5][6].

This aggressive enforcement of export controls follows a voluntary framework established by the Trump administration through a recent executive order[13]. The policy mandates that AI developers submit their most powerful frontier models to the federal government for up to thirty days of evaluation before releasing them to trusted partners[13]. This structured review window is designed to give government scientists and intelligence analysts an opportunity to red-team the systems for potential national security threats. However, the sudden and heavy-handed shutdown of Anthropic’s models has highlighted the tensions inherent in this system, demonstrating that the government is willing to use coercive export control powers when voluntary agreements are deemed insufficient[6][13].

The administration's secretive and aggressive vetting process has drawn sharp criticism from tech industry leaders, venture capitalists, and civil liberties advocates[6]. Critics argue that using export controls to pull a commercial product off the market without public disclosure of the specific technical flaws sets a dangerous and non-transparent precedent[6]. They warn that such unpredictable interventions could stifle domestic innovation, discourage investment in American AI labs, and drive talent to more stable regulatory jurisdictions abroad. Furthermore, some industry insiders argue that the government’s heavy-handed actions might backfire by creating a false sense of security while failing to address the underlying reality that open-source alternatives and foreign models continue to advance rapidly without similar constraints[5].

As the final details of the agreement between Anthropic and the federal government are negotiated, the imminent return of Fable 5 could bring a swift conclusion to an unprecedented stand-off[2][4]. If the model is successfully restored to the public, it will provide a crucial template for how AI labs and government regulators can navigate future security disputes. However, the episode has permanently altered the landscape of the tech industry, establishing a clear precedent that the government will not hesitate to intervene directly in the release of commercial AI[6][2]. The true test for the industry will be whether this emerging co-regulatory model can successfully balance the urgent demands of national security with the rapid, open innovation that has made the American tech sector a global leader[5].