Document arrives
A PDF, a transcript, a web clip, an email. Anything entering your Vault.
Membrane checks
Pattern detection, normalization, sanitization. All before the document reaches your AI.
Clean. Joins your Vault.
Flagged. Never indexed.
What it catches
The attacks most AI tools miss.
Every document is also a potential weapon. The Membrane assumes any input could be hostile and treats it accordingly.
Prompt injection
Hostile text embedded in documents that tries to hijack your AI's behavior.
Jailbreak attempts
Manipulative framing that tries to remove guardrails or redefine the AI's role.
Embedded payloads
HTML, JavaScript, or markup that could leak data when rendered or summarized.
Hidden text tricks
Invisible content designed to manipulate the AI without the human reader noticing.
How it works
Four checks. Every document.
Normalize
Unicode is canonicalized, hidden characters revealed, zero-width chars stripped, control characters removed.
Detect
Pattern matching for known injection signatures, plus learned classifiers tuned on adversarial datasets.
Fence
Untrusted content is wrapped in unforgeable markers so the AI can never confuse it with instructions.
Decide
Clear content is indexed. Suspicious content is quarantined for your review. The AI never sees the raw payload.
Quarantined doesn't mean deleted.
Flagged documents are isolated, marked, and surfaced for your review. You decide what to do. The default is conservative: when in doubt, keep it out of your AI's context.
Inspect
See exactly what was flagged and why
Override
Approve safe content the filter caught
Discard
Permanently remove confirmed threats