This bundle stores translations that every page of an application renders, sends source strings to third-party AI providers with the application's API keys, and imports files. A flaw here reaches every visitor of the host application, so please report it privately.

Do not open a public issue. Use either of the two private channels:

• GitHub security advisory ("Report a vulnerability")
• email to sylius@groupe-cyllene.com

Please include the bundle version, the Symfony and PHP versions, the database, what an attacker could do, and the smallest reproduction you have. Redact every API key from the logs and configuration you attach.

First response within 5 working days. We keep you posted on the analysis, then on the fix and its release date.

Three behaviours are documented rather than defended, because they depend on how the host uses the bundle (where it renders a value, what it sends to a provider, what it approves):

• Translations are untrusted text. A value can come from an AI provider or an imported file. PlaceholderConsistencyChecker compares placeholders; it does not sanitise markup and never rejects it. Escaping at render time stays the application's job: never emit a translated value through |raw, into an HTML attribute or a JavaScript context without the proper escaping.
• Source strings reach a third party. They are sent to the configured provider with the application's API key. Choosing which catalogues may be translated, and by whom, is a deployment decision.
• A source string can steer the model. It is part of the prompt: a crafted value (imported, or typed in a back office) can talk the model into altering the other strings of its batch and reporting a high confidence for them. The prompt tells the model to treat the strings as data, which lowers the odds, not the possibility. The bundle never applies a suggestion without a human decision; an integration must not approve on the self-reported confidence either.

Anything else (a way to read or write overrides without going through the host's own authorisation, a way to make the bundle execute imported content, an injection through a catalogue identifier or a scope code) is in scope, and we want to hear about it.

Coordinated disclosure: the fix is released first, then the advisory. We are happy to credit the reporter, unless they ask otherwise.