The Fable 5 Affair: When The State Pulls The Plug
In June 2026, the US government forced Anthropic to disable its newest frontier version within three days of launch. Read as a story about products, it looks like a recall. Read as a story about systems and human beings, it looks like something considerably stranger. This piece is the first in a series of blog posts by Eleonora Bolsi exploring AI from a philosophical perspective. Eleonora is a researcher in Philosophy of Mind and studied at the University of Malta: her interests run from intelligence in living beings to human relationships with AI.
Anthropic has long sat at the centre of debates on AI, Big Tech, government, and security. Before the ethical implications, the sequence of events must be recalled.
In March 2026, the US Department of Defence classified Anthropic as a “supply chain risk” after the company refused to release Claude for mass domestic surveillance and autonomous weapon systems without ethical restrictions. This frames what happened after April, when Anthropic announced Mythos and an imminent public listing. US pressure and the future company’s Nasdaq entry had already softened CEO Dario Amodei’s ethical stance on the importance of interpretability.
On 9 June 2026, Anthropic released Claude Fable 5 and Mythos 5: the first is a general version presented as a usable, safeguarded Mythos-class system.
Mythos had already been tested after the Preview phase, most visibly through Project Glasswing, where selected companies explored its cybersecurity capabilities. Both models were powerful enough for long-horizon coding, research, and agentic work, but constrained in sensitive domains such as cybersecurity and biology.
Three days later, Fable 5 was gone.
On 12 June, US Commerce Secretary Howard Lutnick transmitted an export-control directive to Amodei, citing national security and prohibiting access by any foreign national, including foreign-national employees. The trigger seemed to be a jailbreak claim by Pliny the Liberator, who had allegedly defeated Fable 5’s classifiers through multi-agent attacks, Unicode substitutions, and narrative framing. Anthropic described the jailbreak as narrow, adding that the government had supplied no technical details beyond security concerns.
Because Anthropic had no real-time mechanism to filter users by nationality, selective compliance was impossible, and Fable 5 has been disabled globally. This was the first government-forced takedown of a publicly deployed frontier AI model in history. It can be read along three lines: AI reification, human agency, and digital sovereignty.
The border control of a Cloud environment
The Fable 5 affair reveals a category error that has structured public reasoning about generative AI from the outset: mistaking a networked capability for a bounded object. The US directive treated Fable 5 as a dangerous artefact whose movement could be policed as a weapons system or a high-end semiconductor, except that the unit of containment is not geographic.
A cloud-based AI system is distributed across infrastructures, customers, jurisdictions, data centres, APIs, employees, and derivative workflows. It is not located anywhere in the sense the directive presupposed, and its users cannot be partitioned along the lines it invoked.
In Cloud Ethics, Louise Amoore argues that algorithmic systems generate forms of action that exceed inherited categories of containment. Ordering the exclusion of foreign nationals from a globally deployed model was to demand a capability no provider possessed. Given that similar capability remained available elsewhere, the directive looks less like containment than imaginary sovereignty. The more powerful the model, the less plausible it becomes that security can be achieved by drawing a line around approved users. The affair warns against a governance imagination oscillating between two pictures: AI as a tool under command, and as a dangerous object to be locked away.
The restriction of human agency
The second line concerns agency.
The official rationale was the danger of unethical or hostile use. Yet this shifts from containing a system to containing human operability around it. The failure to contain AI risk translates into a restriction on human action. Users lost access. Non-US employees were potentially excluded from systems they had helped build.
This is a revealing inversion of human-centredness: the human appears not as the centre of the AI universe, but as the problem.
Because the algorithmic system cannot be rendered fully governable, the implicit message is stark: since we cannot entirely contain AI, we must contain human agency. The algorithmic black box becomes a political one, answered by chains of delegation: citizens depend on firms; firms depend on governments; governments invoke security at the cost of autonomy.
The reported bypass of Fable 5 reinforces the point. Cyber capability emerges from model, tooling, orchestration, user intention, infrastructure, and organisational context. Governing a parameter as a perimeter creates an appearance of control while leaving the conditions of misuse intact: jailbreaks travel as easily as the models they target, and a restriction binding the compliant rarely binds the determined.
As Mark Ryan has argued in his Foucauldian critique of human-centred AI (HCAI), the language of “putting humans at the centre” obscures which humanity is at stake. The Fable 5 sequence supplies an answer: users are constrained, foreign-national employees are constrained, the company is constrained, and constraint flows in one direction. Humanity is invoked as a conceptual label rather than treated as a stakeholder.
Digital sovereignty confined
The third line is digital sovereignty. John Cheney-Lippold’s jus algoritmi described a regime in which identity, entitlement, and exposure were assigned by algorithmic inference rather than classical legal categories. The Fable 5 affair suggests a further displacement: decision shifts from the citizen to the algorithm, and from the algorithm to the state acting on the platform. Call it jus gubernandi: a regime in which governments do not merely regulate AI, but conscript the firms that build it, extending security imperatives into systems themselves.
The paradox is that both the corporate and the governmental versions of sovereignty sideline the human horizon they claim to protect. The framing of an AI system as a containable object reveals the short-sightedness of the proposal across all three lines examined: the impossibility of reifying a cloud-distributed technology, the limits of containing human agency, and the brittleness of sovereignty so conceived. What remains is a contest between institutional actors against the backdrop of a conservative approach to power that excludes human operability as a peer counter-power.
In practice, a jailbreak cannot be confined any more than the model itself. Yet the arm wrestling between governments and Big Tech reveals the underlying intent: to ration human agency and its horizon of sovereignty. A genuinely human-centred approach must preserve human authority in the infrastructures through which AI operates. Otherwise, human-centred AI becomes a label for a world in which humans are centred only as risks to be managed.
Further insights
For more on this topic, watch The Day a Government Switched Off an AI Model.
