
Argent with Presidio, built in
Every Argent gateway ships with Presidio, the open-source engine for finding and removing personal data. It reads each prompt before the model does and each reply before your user does, and takes out what your policy names. Nothing to buy or install.
What Argent does with Presidio
People paste customer records, account numbers and keys into AI tools. Argent takes them out before the text leaves.
Harmonic Security found that 8.5% of the prompts it studied sent to AI tools contained sensitive information (Harmonic Security, 2025). OWASP lists sensitive information disclosure as LLM02 in its Top 10 for LLM Applications 2025 (OWASP GenAI Security Project).
Presidio is an open-source framework for detecting, redacting, masking and anonymizing personal data (Presidio docs). Argent runs it as two services: the analyzer finds, the anonymizer rewrites.
- Inside your deployment. Both services run beside the gateway and are reached over local Unix sockets, so text is never sent to anyone else to be scanned.
- Both directions. Prompts are scanned before they reach the model. Replies are scanned before your user or agent sees them, because models echo and recall sensitive data.
- Pinned and honest. The Presidio release is pinned. Argent refuses a policy that lists a protection the pinned recognizer can never trigger, so a setting on screen is a setting that works.
- A signed record without the value. Each finding is recorded by type and count, never the value itself. A blocked request gets a real refusal that names the types found and the trace id.
What it finds
Presidio's recognizers, plus detectors Argent adds for the data Presidio does not cover well.
| Detector | What it catches |
|---|---|
| Presidio recognizers | Names, Social Security numbers, card numbers, email, phone, IP address, IBAN, bank numbers, passports, driver licenses, medical licenses, crypto wallets and more |
| Secrets pack | About 60 credential rules: cloud keys (AWS, Google Cloud, Azure), GitHub and GitLab tokens, Slack, Stripe, AI provider keys, private keys and more |
| Checksum validators | US ABA routing numbers and SWIFT/BIC codes, checked by checksum plus nearby words |
| Your own patterns | Employee IDs, project codes or any identifier you define, each tested on sample text before you save it |
| Exact Data Match | Your real customer or employee records, matched from keyed digests you upload. Argent never stores the values readably |
- Reads the right text. Argent knows each provider's wire format (OpenAI, Anthropic, Gemini, Bedrock and others) and scans what the model will actually see, not timestamps or JSON structure.
- Looks inside encoded text. When Argent scans a raw body, base64 values are decoded and scanned too, and a request whose decoded content still holds findings is refused.
What happens to a match
You choose, per type of data, what happens when Argent finds it.
| Choice | What the model receives |
|---|---|
| Redact | A typed marker, such as <US_SSN_REDACTED> |
| Mask | The same length, with the characters replaced by * |
| Hash | A stable pseudonym made with a per-tenant key, the same across turns, so the model can still tell two people apart |
| Block | Nothing: the request is refused |
| Restore | A token, and the original is put back into the answer your user sees. The model never saw the value |
On the reply side, the leak mode decides what happens when a model sends sensitive data back: redact it out of the reply, or replace the whole reply with a safe error. Every time Argent applies a weaker strategy than the one asked for, such as hash with no key configured, the signed record says so.
What you configure
Presidio is part of every gateway, in Docker Compose and in the Helm chart. What you set is the policy.
- Start from a safe default. A tenant with no policy is never unprotected: it gets a high-recall default that redacts and records Social Security and card numbers, email, phone, IP, IBAN, bank numbers, passports, licenses and crypto wallets, and counts likely secrets for review.
- Choose what to protect. On the DLP page in the console, pick the types of data, a score floor for each and what happens to each. Choose the scan scope: the whole request, only what users wrote, or off, which is itself recorded.
- Add your detectors. Add your own patterns with a dry run, set the secrets pack to enforce or observe, and upload Exact Data Match sets.
- Decide on replies. Set the leak mode for replies, and turn restore on if your users need their own values back in answers.
- Watch first. In watch mode, DLP works out exactly what it would redact and records it for review, while requests pass unchanged. Then switch to enforce.
Every policy change is versioned and written as a signed event before it takes effect, so who changed what, and when, is always on the record.
Built to fail safe
A scan that did not run is never reported as clean.
- No answer, no pass. If the analyzer or any detector fails, the request is refused with its own reason, never forwarded as if nothing was found.
- No gap if a service is down. If the separate DLP service is not running or not healthy, the gateway keeps scanning with its own built-in copy.
- Fast on long conversations. Long bodies are split and scanned in parallel, and results are cached, so a coding tool that resends the whole conversation each turn pays only for what is new.
- Replies fail closed. If the reply scan faults, the reply is replaced by default. Passing it instead is an explicit deployment setting, and each pass is noted on the record.
- Streams scanned whole. Streamed replies (server-sent events) are put back together before scanning, so a number split across two chunks is still caught.
Where this fits
Pages that go deeper.
- Capabilities: Sensitive Data Redaction, Response Screening, Watch Before You Enforce, Session Replay and all capabilities.
- Goals: data loss prevention, compliance evidence, sovereign AI and all goals.
- Use cases: governing customer-facing AI assistants, rolling out copilots and chat safely, securing AI coding agents, controlling agent tool access and AI in regulated and air-gapped environments.
- Where it runs: air-gapped on-premises, private cloud or your public cloud VPC. Compare them on deployment options.
- Industries: Financial Services, Healthcare, Insurance, Public Sector, Retail and all industries.
- More integrations: every system Argent connects to, grouped by type, on the integrations page.
Want this running on your own use cases?
All downloads