
One prompt, one number, no way back
An AI agent pulls a customer's details into its request so the model can help. If a Social Security number is in that text, it goes to an outside AI provider, and nobody decided to send it.
- In one study of employee prompts to AI tools during 2024, 8.5% contained sensitive information.¹
- IBM found that unapproved ("shadow") AI was involved in 43% of the breaches it studied, up from 20% a year earlier.²
- Data-loss and web-proxy tools built for people and files usually never see an agent's request.
Once an outside provider holds a customer's number, you cannot pull it back.
What Argent does before anything leaves
Argent Strategic Intelligence sits between your AI agents and the models they call. Its data loss prevention (DLP) check reads each request routed through it and replaces sensitive values with a redaction marker before the request leaves.
- The model receives
<REDACTED>where the number was. The rest of the request goes through, and the agent still gets its answer. - Detection runs inside your own Argent deployment, not in someone else's cloud.
- With data protection in Enforce, the default policy redacts from the first request, before anyone writes a policy.
- If Argent cannot finish the check, it refuses the request rather than send it.
The model's reply is screened by a separate control: Response Screening.
How it works, step by step
- The agent sends its request through Argent. Any agent or tool with a base-URL or proxy setting can be pointed at the Argent gateway.
- Argent reads what the person authored. By default it scans what was typed or pasted and what the agent's tools brought back, in every turn. The system prompt and tool definitions are left alone; you can switch to scanning the whole request if you want it.
- It recognizes the data by shape and context. A dashed Social Security number is caught on its shape alone, and scores higher when words like "social security" sit nearby. Card numbers and IBANs are checked against their checksums.
- It applies your policy. Under the default, Redact + Audit, the value is swapped for a marker and the model provider receives only the redacted version.
- It writes a signed record. The ledger line reads like "DLP redacted SSN x1": the kind of data and how many, never the value.
- It refuses when it cannot check. If the detector is down or a request is too large to scan, the request is refused, not sent.
What the operator sees
Everything an administrator needs to set the policy and to prove what happened, without the data itself spreading.
- The DLP page. Scan scope, what happens on detect, the replacement style, confidence thresholds and which kinds of data are on. Only a SuperAdmin or the tenant's own TenantAdmin can change it, and every change is a signed event.
- Two copies, side by side. On the Replay page, an authorized administrator can open a request and see "Sent by the agent (raw, before DLP)" beside "Forwarded by Argent (after DLP)", with the marker highlighted.
- A recorded reveal. The raw copy stays hidden until the reader clicks "Reveal what the agent sent" and confirms "This read is recorded in the signed audit log under my name." If that record cannot be written, the text is not shown.
- Counts by data type. The home screen shows what DLP caught by kind, never the values.
Go deeper with Session Replay and One-Click Audit Evidence.
Redact, record or refuse: your call
| Setting | What happens |
|---|---|
| Redact + Audit (default) | The value is replaced with a marker, the request goes on, and the record keeps the kind and count. Masking with asterisks is also available. |
| Audit Only | The request is sent unchanged, and the record says what was detected. |
| Deny | The request is refused with a plain message such as Argent DLP blocked the request: US_SSN detected (trace ...). Nothing is forwarded, and the message never contains the value. |
| Watch mode | Argent works out the redaction, forwards the original, and records what it would have done. |
What it looks for
- On by default: Social Security, card, bank account, IBAN, passport, driver license and medical license numbers; email addresses, phone numbers, IP addresses and crypto wallet addresses; secrets such as API keys are counted.
- Switch on more: person names, places, dates and web addresses; a pack of about 60 credential rules (cloud keys, AI-provider keys, private keys); US routing numbers, checked by checksum, and SWIFT codes, checked by format and a banking word nearby.
- Your own formats: an admin writes a custom pattern, tests it on sample text and sees the matches before it goes live. Every pattern change is signed before it takes effect.
- Your own records: Exact Data Match fingerprints your real customer or employee lists on your side; only scrambled, one-way fingerprints are uploaded, never the values.
- Your own engine: if you run your own Microsoft Presidio engine, Argent can be set up to send the check to it instead, and refuses the request if that engine fails.
Who it serves
Any team whose AI agents handle customer, employee or account data. Start from your goal, your use case or your industry.
- Goals: Data loss prevention, Compliance evidence, Audit and accountability, Sovereign AI and all goals.
- Use cases: Governing customer-facing AI assistants, Rolling out copilots and chat safely, Securing AI coding agents and AI in regulated and air-gapped environments.
- Industries: Financial Services, Insurance, Healthcare, Retail, Public Sector and all industries.
For financial firms, the pressure is already written down.
- FTC Safeguards Rule, 16 CFR 314.4 (non-bank financial institutions): a security program with access to customer information for authorized users only, encryption, monitoring, and oversight of service providers.³
- New York DFS, 23 NYCRR 500.17 (covered entities): notice "as promptly as possible but in no event later than 72 hours after determining that a cybersecurity incident has occurred at the covered entity, its affiliates, or a third-party service provider."⁴
How a team starts
Watch first. Nothing changes for staff until you decide.
- Watch. Argent records what it would have redacted while requests pass unchanged. See Watch Before You Enforce.
- Enforce. Switch data protection to Enforce; the default policy redacts from the next request.
- Tune. Add your own patterns, switch the policy to Deny if any detected data must stop the request, and review decisions in Replay.
- Inside your walls. Argent runs in your own Kubernetes cluster, as a single-node bundle, or air-gapped on premises. Compare deployment options. Redacted prompts go on to the model provider you chose.
See the whole platform, take the walkthrough, or browse all capabilities.