The Claude Fable logo is showcased on a smartphone screen (Image by Samuel Boivin/NurPhoto via Getty Images).
In June, the public was given a brief opportunity to interact with a model named Fable 5, but within three days, the federal government stepped in and restricted access to it.
While Fable 5 wasn't Anthropic's most advanced model—that title went to Mythos 5, which remains under tight control—Anthropic classified Fable 5 as "Mythos-class," claiming it surpassed all prior models available to general users, especially in software engineering and scientific research. The company expressed concerns that these capabilities could be repurposed for harmful use if not properly regulated.
Just three days post-launch, citing national security reasons, the government imposed export controls that prohibited access to anyone classified as a foreign national, including Anthropic’s own international staff. Unable to verify compliance on an individual user basis, Anthropic made the decision to revoke access entirely.
The introduction of Mythos and Fable signifies an ongoing challenge for government oversight. The rapid pace of AI development has outstripped regulatory frameworks, with new functionalities emerging faster than authorities can manage.
A recent paper from Google DeepMind entitled "From AGI to ASI" explores this trajectory, mapping the evolution from artificial general intelligence—historically the primary goal of the field—to artificial superintelligence, a stage where machines surpass human capabilities. Although such conversations were previously theoretical, the limited release of Fable 5 has provided a material reference point.
Governments engage in lengthy deliberations, while the AI sector operates at breakneck speed, with advancements occurring in mere weeks as opposed to fiscal years often relied upon by regulators.
Reports indicate that the White House and Anthropic are currently formulating a system to assess the severity of security flaws in new models and determine when intervention is necessary. However, such standards were not in place during the Fable 5 incident, and they are still being established.
The DeepMind paper identifies four potential pathways for the evolution from AGI to superintelligence, which can occur simultaneously: scaling up existing models with more data and computing power, genuine algorithmic breakthroughs, recursive self-enhancement by AI performing research, and large networks of cooperative agents. Any of these processes could already be in motion.
This last scenario poses a significant challenge to current regulation models, which often wait until a technology matures before assessing its impacts. AI, however, is in a constant state of flux, often serving as a catalyst for its future iterations.
The paper’s authors present a measured perspective on the concept of superintelligence, acknowledging obstacles such as insufficient data, rising resource demands, and the intrinsic complexities of cutting-edge research. They also recognize the potential for regulatory and public pushback as inhibiting factors. Yet, such friction may only constrain technological progress if regulatory bodies can keep pace with rapid advancements; the Fable 5 case suggests they are not yet able to do so.
Unanswered questions persist: When does providing API access equate to an export? A vulnerability that seems trivial one week might transform into a national security concern the next, with no consensus on where to draw the line. Who determines whether a model has excessive capabilities, and what criteria—compute power, operational use, or context—guide these assessments?
The likelihood of similar incidents reoccurring remains high, and future challenges may extend beyond cybersecurity. They could involve models with the ability to engineer a pathogen or manipulate a power grid or execute large-scale influence operations. Such scenarios will demand a level of technical expertise that most regulatory agencies currently lack and invoke legal frameworks that are often vague, paired with the urgent need for international cooperation that typically arrives postpartum.
This may signal a turning point for governments to establish robust evaluation frameworks, such as pre-release assessments. However, the more probable outcome in the near term appears to be reactive measures, as agencies strive to catch up with each new model post-release.
DeepMind’s authors caution that preparing for the aftermath of AGI will necessitate ongoing forecasting, benchmarking, and vigilance, along with the agility to quickly translate findings into actionable policy. They describe this challenge as navigating a "high-velocity technological trajectory," highlighting the limited ability to manage the situation other than closely monitoring developments and adapting in real time.
While Fable 5 will soon fade from public discourse, its debut serves as a crucial example of the emerging dynamics. Within just three days, a product launch escalated into an export control dilemma and a compliance challenge, ultimately involving geopolitical negotiations. The lessons learned from this case underscore the likelihood of future occurrences, with scant expectation for labs to moderate their pace while regulations are still under development.
