Why was Claude Fable 5 banned?

The real reason Anthropic disabled Fable 5 for its users and what it means for product designers

On Friday, June 12, Anthropic received a request from the US government to restrict access to its latest and most powerful AI model, Claude Fable 5. This model was available for general usage only for a couple of days, and now, if you try to use it in Claude Code, you will see the system error message: “Claude Fable 5 is currently unavailable.”

In this article, I want to explain why Anthropic rolled back Fable 5 and what long-term impact it will have on the industry and product designers in particular.

Real reason: Fable 5 is prone to jailbreaks

When Anthropic received the request from the government, officials cited a “national security threat” as the primary reason for the ban. Despite all the investment that Anthropic made in building the most secure guardrails for AI model, Fable 5 was jailbroken within 48 hours of its public release.

A researcher known as Pliny the Liberator publicly claimed to have bypassed those guardrails anyway. The techniques weren’t new. The researcher used a combination of Unicode/homoglyph substitution (replacing standard ASCII characters with visually identical Unicode equivalents), long-context framing (baking the adversarial instruction deep inside a large document or conversation), narrative fiction framing (wrapping the harmful request in a creative fiction context), and a decomposition-recomposition strategy (splitting a single harmful request into multiple seemingly harmless sub-prompts).

The use case for jailbroken Fable 5 was serious: Pliny the Liberator proved Fable 5 could easily be used for attacks on crypto protocols and software vulnerabilities.

Is AI model jailbreak a unique situation?

Not at all. In fact, Dario Amodei, CEO of Anthropic, mentioned that most AI models are prone to jailbreak even before the Fable 5 release. Anthropic invested a lot of effort in making it the most protected AI model to date by baking a lot of guardrails into it, but it was not enough, and the model was jailbroken.

[embed]Claude Code Guardrails Practical guide for creating guardrails for Claude Codeuxplanet.org

Quite possible that situations like this will become more common in the future: AI companies will release powerful new models, discover unexpected risks, and roll them back shortly after launch.

What does it mean for product designers?

Fable 5 jailbreak means that

AI design workflows are no longer just a tool choice; they are infrastructure risk.

So for product designers, the lesson is not “don’t use AI,” but rather “don’t build your entire design process around one closed frontier model.”

1. Your workflow should be model-agnostic

Prompts, product briefs, design-system rules, UX heuristics, and evaluation checklists should work across all three popular AI models: GPT from OpenAI, Gemini from Google, and Claude from Anthropic.

For every AI workflow, know your Plan B: which AI model you will use, and why, if your primary AI model becomes unavailable.

You need fallback model(s). If one model disappears (like Fable 5), gets nerfed, or becomes super expensive (like Opus 4.8), your workflow should still function with another model, such as GPT.

2. Design system should remain the source of truth

AI can generate screens, flows, and copy, but final tokens, components, interaction rules, and brand logic should live outside the AI chat. AI should never be the final authority that you rely on when making your design decisions.

Never rely on AI memory as documentation

If the model “knows” your product only inside one chat, your process is fragile. Move important context into stable files: DESIGN.md, CLAUDE.md, brand guidelines, component docs.

[embed]What is DESIGN.md and How To Use It One of the biggest challenges when using AI design generators is producing consistent output. Even when you feed AI…uxplanet.org

3. Do not paste sensitive product context blindly

As the Fable 5 example taught us, even the most powerful models are prone to jailbreaks. Thus, treat important information, such as billing details/API keys, unreleased features, user research, metrics, and internal strategy, as sensitive by default.

Want to master Claude Code?

Check out my complete guide to Claude Code, packed with highly practical insights on how you can integrate it into your design process. This guide is updated on a weekly basis, so you’ll always get the latest information about Claude Code.

[embed]Claude Code: Practical Guide for Product Designers Practical guide for product designers who want to master Claude CodeIt covers the following topics: Claude Code Best…babich.gumroad.com