The Cobra Effect Comes for AI

The Cobra Effect: How the American Government’s Shutdown of Fable 5 Undermines Security and Liberty

Photo by Steve A Johnson on Unsplash

Written By: Bryan Theunissen

Liberal societies made an unusual wager. They decided, over several centuries and through considerable internal resistance, that the dangers created by free access to knowledge were outweighed by the benefits. The printing press was not controlled. The scientific method was not licensed. Universities were not required to seek government approval before publishing inconvenient findings. Open-source software was not treated as a weapons system. Each of these decisions carried real risk. Each of them also produced the technological and intellectual dominance that characterises the West today. The wager paid off — not despite the openness, but because of it.

The American government’s shutdown of Anthropic’s Fable 5 and Mythos 5 on 12 June 2026 suggests a growing willingness to abandon that bargain. Whether that abandonment will improve security remains entirely unproven. That it weakens a tradition of intellectual liberty built over centuries is already evident.

What Happened

The American Department of Commerce, citing national security authorities, issued an export control directive requiring Anthropic to block access to Fable 5 and Mythos 5 for all foreign nationals — including the company’s own foreign-national employees. Unable to segregate usage effectively, Anthropic disabled both models worldwide within hours of receiving the order at 5:21pm on a Tuesday. The letter provided no specific details of the national security concern.

Fable 5 represented a significant advance in reasoning, coding, and complex task execution, developed in close cooperation with the American government itself, the UK AI Safety Institute, and multiple independent red-team organisations. Thousands of hours of adversarial testing preceded its deployment. It was, by any reasonable measure, the most carefully governed frontier model ever released.

Its termination has all the hallmarks of the Cobra Effect: an intervention designed to solve a problem that instead creates a larger one. The American government set out to contain a security risk. What it actually achieved was disarming the defenders, rewarding the adversaries, and sending a signal to the entire industry that transparency is a liability.

What the American Government Actually Found

Anthropic’s public statement is worth reading carefully, because it reveals the technical basis — or rather, the absence of one — for this decision.

The American government’s stated concern was a method of jailbreaking Fable 5: bypassing its safeguards to elicit information about software vulnerabilities. Specifically, the technique involved asking the model to read a codebase and identify any flaws. Anthropic reviewed the demonstration and reached a straightforward conclusion: the capability on display was already widely available in other publicly deployed models, including OpenAI’s GPT-5.5. It is, in fact, a capability used every day by the security professionals and software engineers who defend systems against attack.

This is not a minor technical quibble. It is the entire argument. If the justification for shutting down Fable 5 is that it can identify software vulnerabilities when asked, and if that same capability is already present in models that remain freely available, then the shutdown achieves precisely nothing in terms of containing the risk. The regulation designed to improve safety has instead selected against transparency. That is an empirical claim, and it is testable.

Selection Pressure Against Openness

The most striking feature of this episode is that Anthropic’s extensive monitoring infrastructure appears to have increased its regulatory exposure rather than reduced it. Anthropic had required 30-day retention of all customer data — a policy that carried real commercial costs — specifically to enable rapid detection and mitigation of jailbreak attempts. A company that invested heavily in visibility, auditing, and disclosure became easier to scrutinise than competitors operating with far less transparency.

Whether intended or not, this creates a selection pressure against openness itself. The industry will observe what happened to the most cooperative, most transparent, most safety-invested actor in the field, and it will draw the rational conclusion: build in private, engage minimally with government oversight, publish nothing that can later be used against you. The collective safety research ecosystem, already fragile, takes a serious blow — not as a side effect of the policy, but as its foreseeable consequence.

This is the Cobra Effect at its most precise. The American government has made the field less safe by punishing the behaviour that makes it safer.

Knowledge, Power, and the Limits of Export Control

Underlying this episode is a conceptual error with deep implications. The American government is treating trained model weights — patterns of mathematics encoding human knowledge — as though they were physical armaments subject to export control. The closer AI models come to functioning as general-purpose cognitive tools, however, the weaker the analogy to controlled weapons systems becomes, and the stronger the analogy to protected forms of knowledge and expression.

This matters because free societies have historically drawn a firm distinction between restricting actions and restricting ideas. Export controls on weapons make sense because weapons are instruments of force. Export controls on knowledge are categorically different, because knowledge — once it exists — diffuses across borders regardless of edicts. Heavy restrictions primarily penalise rule-following actors while leaving determined adversaries entirely unaffected. The American government has not prevented the capability it was concerned about from existing in the world. It has simply removed the most safety-conscious implementation of that capability from the field.

State actors and laboratories unbound by Western democratic norms face no equivalent constraints. They can now pursue matching or superior capabilities without restriction, without Anthropic’s ethical guardrails, without its alignment investments, without its monitoring architecture. American defenders, meanwhile, have been handed older tools. The directive designed to protect American security interests has exported the capability advantage and imported the vulnerability.

The Precedent

The standard applied here — that a narrow, non-universal jailbreak producing capabilities already freely available in competing models is sufficient grounds for global shutdown — would, as Anthropic correctly notes, effectively halt all new model deployments across the entire frontier AI industry. If that is the intended policy outcome, it should be stated openly and debated democratically. If it is not, the directive reflects a level of technical illiteracy that is itself a security risk.

The fallout is already visible. Talent and capital are accelerating moves to jurisdictions with more predictable regulatory environments — the UAE, Singapore, and others that offer robust infrastructure without the threat of arbitrary shutdown. This is not speculation about future risk. It is the rational response of rational actors to a demonstrated threat. Free societies became technologically dominant because they tolerated the risks associated with open knowledge. A policy that drives frontier AI development into less transparent, less scrutinised, less safety-conscious environments does not reduce those risks. It concentrates them, offshore, beyond the reach of the very oversight apparatus the American government claims to be protecting.

Principles Over Panic

A rational approach would recognise that technological leadership, transparency, and aligned innovation are America’s strongest assets in this domain — not liabilities to be regulated away. Targeted safeguards, international norms where achievable, and continued investment in safety research offer far better paths than blunt prohibitions that cannot be enforced globally and that punish precisely the actors most inclined to cooperate.

The danger revealed by the Fable 5 shutdown is not that American policymakers are malicious. It is that complex technological ecosystems respond to incentives rather than intentions. A policy can be enacted in the name of security and still reduce security. It can be enacted in the name of safety and still discourage safe behaviour. It can be enacted in the name of national advantage and still accelerate the migration of talent, capital, and innovation elsewhere. The Cobra Effect is not a failure of motive. It is a failure of understanding. And in a field advancing as rapidly as artificial intelligence, that distinction matters enormously.

The wager that built Western technological civilisation was a bet on openness. On 12 June 2026, the American government cashed it in for an afternoon’s worth of security theatre.

Bryan Theunissen is a South African doctor with a stubborn streak of optimism.