Grodmir — AI gateway protection limits
Grodmir reduces personal data, secrets and confidential material reaching model providers; this AI gateway does not make leaks impossible. The full protection limits follow.
Grodmir is an AI gateway for data protection and AI usage governance in organisations.
Which traffic and files does Grodmir not cover?
- - Traffic that does not pass through the gateway, the extension or the MCP proxy: personal provider keys, other browsers, unproxied MCP servers, tools without a base-URL setting. Block provider hosts in your egress firewall. Grodmir reports uncovered channels where it can see them (extension coverage; aiguard mcp scan for MCP servers configured on a workstation, locally only) but cannot see the rest.
- - Images and files (PDF, DOCX, XLSX, images): not analysed unless the policy switches on defaults.attachment_inspection (text, or text_and_ocr with the separate aiguard-dataplane-ocr image). Without it, and for anything that cannot be read (scanned PDFs, encrypted or damaged files, low OCR confidence, URLs or provider file ids), the policy decides; the default is to block. A PDF with pictures on its pages counts as unreadable too, because the pictures are not read. A PDF, office file or image is never rewritten: if its text would need pseudonymisation or removal, the request is blocked. OCR quality depends on the scan and is not guaranteed.
- - The optional classifier for health data, salaries, contract terms and trade secrets is a simple keyword lexicon: it only marks text (approval rules on it work in shadow mode only), misses paraphrases and has not been measured on real documents.
- - Voice, and languages other than Polish, German and English.
Where can Grodmir detection miss data or raise false alarms?
- - Identifiers with checksums (PESEL, NIP, REGON, IBAN, cards, German tax IDs) and secrets are detected by rules and are measured precisely on synthetic data.
- - Names, organisations and addresses come from a small local model trained only on synthetic data. It misses entities and raises false alarms; quality on real documents is not validated. On three separate hand-written synthetic sets it covered 24 of 35, 73 of 120 and 64 of 120 entities. Measure it on your own data with aiguard eval before you rely on it. Use dictionaries for terms that matter, and run observe mode first.
- - Deliberate obfuscation by the person sending the text is not reliably caught: encodings (base64), homoglyphs, a value split over several fields, or digits separated by spaces (4 4 0 5 …). Invisible characters (zero-width, soft hyphen, word joiner) and full-width digits are normalised and do not hide a value. Grodmir protects against mistakes, not against a user determined to exfiltrate data.
- - Numbers without context words may fall below the confidence threshold.
- - Confidentiality markers (doc.confidential_marker) are labels: "POUFNE", "Vertraulich" on its own line, "Strictly confidential", "Tajemnica przedsiębiorstwa", "Klassifizierung: vertraulich". The word inside a sentence ("this note is confidential") is not a marker.
- - Internal network names (network.internal): RFC 1918, 100.64/10, link-local and IPv6 unique-local addresses, and host names ending in .internal, .intra, .intranet, .lan, .corp, .home.arpa, .localdomain or (with two labels) .local. Other internal names need a dictionary.
- - Passwords are found as key=value / key: value and in sentences ("my password is …", "Mein Passwort lautet …", "hasło to …") when the value contains a digit or symbol. A password without such a phrase is not recognised.
- - Dictionary phrases match any whitespace, any dash variant (or none) and Markdown emphasis between words, but not look-alike letters from other alphabets.
Which behaviour limits apply to pseudonyms and observe mode?
- - Model thinking (Anthropic thinking blocks) keeps its pseudonyms in the answer: those blocks are signed and sent back unchanged, so restored values would reach the provider again.
- - Pseudonyms are restored for the user, not in tool-call arguments (unless a tool is on the allow list) and not across MCP tool results (Q-044).
- - Observe mode never applies policy blocks or changes content, including when a detector fails. Authentication, admission, routing and unavailable inspection services still refuse traffic. Browser/MCP inspection cannot execute a route and refuses it in enforce mode.
- - A local administrator with full control of a host can switch the gateway off; the audit trail shows it.
- - Chrome may block the browser extension from reaching a private-address gateway (Q-043); the extension then fails closed.
- - The extension's page selectors have only been tested against look-alike pages (Q-042).
What has not been independently verified yet?
- - No external penetration test has been completed (Q-050). Do not treat the product as audited.
What do attestations prove, and what do they not prove?
- - Schema 2 adds hashes of published policies, destinations with their historical regions, and applicable published exceptions for the client and reporting period. No policy text or dictionary terms are copied. Existing schema 1 signatures remain verifiable.
- - Current telemetry has no executed policy version or exception-use identifier. Publications therefore do not prove device adoption or actual use of each exception; local audit remains the source for request-level evidence. Regions list possible published values during the period; an empty list means historical region evidence is unavailable.
- - The extension covers paste/drop, not typed text or other upload mechanisms. MCP server stderr is suppressed because arbitrary server diagnostics may contain sensitive content.
- Grodmir — Prompt DLP for ChatGPT in the browser
- Grodmir — Cursor AI gateway setup and coverage limits
- Grodmir — MCP proxy for spreadsheets
- What do the Grodmir categories mean?
- Grodmir — AI gateway selection criteria
Content updated:
Categories: AI gateway · Prompt DLP · MCP proxy · AI usage governance