Fable 5 shutdown reveals a dangerous gap between AI regulation and AI reality
Follow Us
As early as June 12, the US government issued an export control directive for Anthropic to halt all access to their Fable 5 and Mythos 5 models due to a newly found jailbreak that posed a threat to national security. The company complied but expressed disagreement, and that was where it got interesting.
SurveyAlso read: Switched to Claude from ChatGPT? Fix these 3 privacy settings right now
The particular jailbreak in question requires one to request the model to analyze a software codebase and find vulnerabilities. Anthropic’s investigation has shown that it could be done quite easily with any GPT-5.5 model as well, including the use of the feature on a regular basis in cybersecurity practices. In other words, the US government halted two highly capable models for hundreds of millions of people because a task like that can be done in any frontier model.
This is what made the situation truly problematic. Anthropic had openly declared, very clearly, that perfection against jailbreaks cannot be achieved anywhere in the industry. This does not represent a failure. On the contrary, this is a realistic reflection of complex systems. Safety is not an absolute yes-no in engineering. You do not secure it forever once you have it. You minimize exposure to danger, stack defenses, detect any breach, and react to it quickly. This is exactly what Anthropic had implemented in its Fable 5 system, even at the price of requiring data retention for 30 days to ensure prompt jailbreak detection.
Also read: Buying a robot vacuum cleaner with mopping under Rs 50000? Check these key features first
None of the above considerations came into play. The fact of a specific jailbreak of non-global application was known to a government official. This was sufficient for a product recall impacting all customers on the platform. There was not a hint of any actual harm caused. There was no evidence that the jailbreak produces harmful output. It did not matter. By Anthropic’s own admission, the jailbreak requires no access specific to Fable to be performed.
It is not the difference between AI firms and regulators that creates such a gap. It is a profound distinction between two concepts of what safety is. As the regulator works in the environment of clear criteria (does the chemical plant comply with the regulations on emissions? Has the drug passed all clinical trials?), these criteria are applied to an industry where there are no limits and threats evolve daily. The consequence is a set of rules, adherence to which would mean grounding all frontier models. Anthropic stated this in their announcement.
The problem I am worried about is inconsistency. While OpenAI’s GPT-5.5 is capable of doing what was thought to be prohibited in relation to Fable 5, it is still accessible. The issue here is neither capacity nor the threat. The distinction is that one company had their product flagged first.
Governments need some means of controlling frontier AI. This is obvious. What is not a regulation but manipulation is oversight that relies on rules unattainable by any developer and is exercised arbitrarily, while the specifics of the issue remain confidential for the developers.
Also read: Asus ROG celebrates 20 years: Pro Art PZ14, Zephyrus Duo come to India
