Claude Fable 5.2: why ownership wording and fallbacks matter
Key takeaways
- One sentence in a prompt, telling Claude Fable 5.2 the codebase is its own, roughly raised how often it merged repeated code from under a tenth of tries to about half
- Fable 5.2 hands flagged requests to older, weaker models, and agent attack success rose from 1.02% to 4.09% on one benchmark
- Teams letting agents read inboxes, websites or third-party repos can switch off automatic model switching, at the cost of convenience
How should a team prompt Claude Fable 5.2, and what should it change in its settings before pointing the model at untrusted content? Anthropic's documentation and its system card, both published around the model's 9 November launch, answer two separate questions that turn out to be linked: how the model treats code, and what happens when it quietly hands a request to a different model.
What Claude Fable 5.2 actually is
Claude Fable 5.2 is Anthropic's generally available model released on 9 November. Notebookcheck reports that it generates output up to 30% faster than Fable 5.1 and that early testing showed about 55% fewer blocked requests. It is the same underlying model as Claude Mythos 5.2, which Anthropic offers only to vetted organisations. The Hacker News reports Anthropic describing Mythos 5.2 as its most capable model for cybersecurity engineering, with access through its Cyber Verification Program.
To make Fable available to everyone, Anthropic runs classifiers, automated filters that check every request for offensive cybersecurity, risky biology and parts of AI development. A flagged request is answered by a less capable model instead: Claude Opus 4.8 for cyber topics, Claude Opus 5 for biology. Claude shows a notice and labels the response with the model that answered.
How prompt wording changes the model's behaviour
Anthropic's prompting guide, quoted by Simon Willison, says Fable 5.2 responds closely to what a prompt says about who owns the code. If a prompt says nothing about ownership, or presents the codebase as someone else's, the model tended in small coding tests to add to the existing structure, even when that repeated existing code.
Think of a contractor in a rented flat versus an owner. The contractor bolts a new shelf onto the wall as it is; the owner moves the wall. The suggested fix is one sentence: "This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it." In Anthropic's testing, the model merged repeated code in about half of tries with that sentence and in under a tenth without it.
How fallbacks work, and where the risk comes in
The detail that matters for security is on Anthropic's help page, according to Notebookcheck: the checks review everything the model reads, including memory, connector content, web search results and files. Text the user never typed can therefore trigger a switch.
That matters because of prompt injection, an attack in which someone hides instructions in an email, web page or file, hoping an agent follows them as if the user had asked. The system card measured the effect. On Gray Swan's indirect prompt injection benchmark, attack success after 15 attempts was 4.09% against Fable 5.2, 1.02% against Fable 5.1 and 0.96% against Opus 5.5. Of the attempts against Fable 5.2, 53% were served by Opus 4.8, against 23% for Fable 5.1. Attacks succeeded 2.05% of the time when Fable 5.2 answered itself and 5.17% when Opus 4.8 did.
A second test used an adaptive attacker with 200 attempts per coding scenario. Without safeguards, 89.82% of attempts against Fable 5.2 succeeded, and almost all of the responses came from Opus 4.8. With the prompt injection probes Anthropic runs in its products, the rate fell to 28.16%, against 11.13% for Opus 5.5 under the same conditions. None of the 43 requests Fable 5.2 answered itself in that test were compromised.
Who uses it and why
The model targets coding and agent work. The Hacker News reports that Fable 5.2 completed 76.3% of challenges on Irregular's CyScenarioBench, against 61.7% for Fable 5.1 and 61% for OpenAI's GPT-6 Cyber. It also scored 89.6% on Anthropic's CVE Rediscovery benchmark, which covers 1,829 advisories for open-source projects published in October 2026. It is available on Amazon Web Services, Google Cloud and Microsoft Azure.
A concrete scenario: a team points Fable 5.2 at a third-party repository to review a dependency. A README contains hidden instructions, and some of the content trips a classifier. The request moves to the weaker model, which in the tests was more than twice as likely to succumb to injection as Fable 5.2 answering itself. Notebookcheck notes that injected content in coding tools is a live risk, citing the tampered Claude Code extension on Cursor's marketplace at the end of October. Related reading: hidden commands found in website headers targeting AI.
What the limitations are
The security numbers come from the system card, so they are vendor-reported. Switching off fallbacks is also not free: a flagged request pauses, and the user must edit the message or pick another model by hand.
What this means in practice
Automatic switching is on by default on the web, in the mobile and desktop apps, in Cowork and in Claude Code. It can be disabled under Settings and Capabilities, or in Claude Code under Config and MODEL & OUTPUT, via the toggle "Switch models when a message is flagged". Teams letting Fable 5.2 work through inboxes, websites or third-party repositories can do so to stop those requests being handed to the weaker model. Anthropic itself recommends starting with Claude Opus 5.5 for most tasks and moving up to Fable only for the most demanding work.
For prompts, the ownership sentence is worth adding when refactoring is wanted, and leaving out when it is not.
On the UK side, the NCSC's guidance on AI and cyber security and its guidelines for secure AI system development are the places UK organisations will look for secure deployment advice. The source gives no UK pricing; it says only that Pro subscribers can use the model within their plan until 16 November, after which it normally runs on usage credits. For context on how quickly these releases arrive, see four frontier models shipped in one week.
Where this is heading
Frequently asked questions
Does telling Claude Fable 5.2 that the code is its own change its behaviour?
In Anthropic's small coding tests, the model merged repeated code in about half of tries with the ownership sentence and in under a tenth without it. Without the sentence, it tended to add to the existing structure.
Why does Claude Fable 5.2 switch to another model?
Classifiers check requests for offensive cybersecurity, risky biology and parts of AI development. When one flags a request, Claude Opus 4.8 answers for cyber topics and Claude Opus 5 for biology. The switch is shown to the user.
How do I turn off automatic model switching in Claude?
Open Settings and Capabilities, or in Claude Code go to Config and MODEL & OUTPUT, and disable "Switch models when a message is flagged". Flagged requests then pause so the user can edit or pick another model.
Is Claude Fable 5.2 more vulnerable to prompt injection?
On Gray Swan's benchmark, attack success after 15 attempts was 4.09% for Fable 5.2 against 1.02% for Fable 5.1. Anthropic attributes the gap mainly to fallbacks to the older Opus 4.8 model.