Anthropic Disabled Fable 5 for More Than an Ordinary Bug: The Model Ran Into National-Security Requirements
Why did Anthropic disable Fable 5 not as an ordinary bug, but as a system that had run into national-security requirements?
Compare Anthropic and Fable-5 on a real task instead of choosing by one benchmark or announcement; the assessment must compare quality, price, access, memory, and control on a real task rather than by one announcement or benchmark.
What to watch for
Key takeaways
An ordinary release is rolled back for bugs or load. Fable 5 shipped, went mass-market, showed real gains — and was switched off together with Mythos 5 three days later. The reason was not the model’s quality but how the state judged its capabilities.
The question is put bluntly: if Fable 5 must be blocked, why not everything else with the same capabilities. The answer decides who commands frontier models — the market, the companies, or the state — and it reaches every business built on AI.
Mythos hunted vulnerabilities in critical software with about fifty partners, and only vetted companies had access. The preview alone found a thousand high-severity issues, including ones in major operating systems and browsers.
The project turned up over ten thousand high- and critical-severity findings, and in open source 90.6 % of the checked ones were confirmed. Yet fixing a single serious vulnerability took about two weeks on average — discovery has outrun repair.
The order required access for foreign nationals to stop. The company has no system that establishes a user’s status — nobody uploads a passport at sign-in. The practical effect: the model went off for everyone, so the directive would not be broken.
The letter does not say what happened, who attacked, or what harm was demonstrated. Nor why the same standard was not applied to others: Anthropic points directly at known problems in GPT-5.5.
Anthropic reviewed the demonstration: the method surfaced a small number of already-known or minor vulnerabilities, and publicly available models can do the same. The company is complying with the directive while fundamentally disagreeing with it.
Prompts and responses are retained for thirty days for trust and safety, and that data can be analysed. Microsoft barred its employees from Fable in the very first days for exactly that reason: the leak risk outweighed the gain in quality.
In its June policy package the company argued that the most powerful models need testing, independent evaluation, and a legal way to stop a dangerous release. The state used exactly that power — against the company itself.
What this episode is about
Fable 5 and Mythos 5 had already reached a mass audience when access was closed for everyone. Reinforced safeguards did not satisfy the US government, leaving Anthropic between users, state requirements, and its own safety principles. This is a new kind of product risk.
An ordinary failed release is rolled back because of errors or load. The Fable 5 story is different: the model was launched, given broad access, shown as an improvement, and then disabled together with Mythos 5. The reason is connected not only to quality but to how the state assesses its capabilities.
Anthropic emphasized that Fable launched with stronger protections than earlier versions. But safety cannot be proved by one internal testing system. The government examines military, cyber, and intelligence scenarios in which even a small increase in capability changes the level of risk.
The company finds itself in a conflict of roles. Users have already built the model into their work and expect stable access. Anthropic has to protect the technology and comply with government requirements. At the same time, it has built a reputation as an organization that places safety ahead of speed. Disabling the model validates the principle while damaging trust in product predictability.
The case shows that the strongest model no longer resembles an ordinary software update. It can be restricted for political reasons, by geography, or by customer type. A business that built a critical process on one provider receives the risk of a sudden shutdown.
The right response is not to declare Fable bad or dangerous because it was disabled once. We need to understand which specific capabilities caused concern, which conditions would restore access, and how the company will compensate users for disruption. AI is becoming a strategic asset, so a release note no longer describes every force governing the product.
We need to understand which specific capabilities caused concern, which conditions would restore access, and how the company will compensate users for disruption. As a result, AI is becoming a strategic asset, so a release note no longer describes every force governing the product.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 4 segments: 2 identified, 0 mixed, 0 probable, and 2 unresolved.
Read transcript on a separate page
Loading…