
If your goal is data loss prevention
The leak path has moved. Customer records, account numbers and keys now leave inside AI requests: typed by a person, pasted from a file, or pulled in by an agent's tools. Then the model can repeat them back.
- In one study of employee prompts to AI tools during 2024, 8.5% contained sensitive information.¹
- IBM found security incidents involving unapproved ("shadow") AI more than doubled, to 43% of the breached organizations it studied, from 20% a year earlier.²
- GitGuardian counted 1,275,105 leaked AI-service secrets in 2025, up 81% in a year.³
- OWASP lists sensitive information disclosure as LLM02 in its 2025 Top 10 for LLM applications: models "risk exposing sensitive data, proprietary algorithms, or confidential details through their output."⁴
Argent Strategic Intelligence sits between your AI agents and the models they call, so it checks each request and each reply on the way through.
Every way data leaves, and what stops it
Data loss prevention for AI is not one control. Here is each path out, and the Argent capability that covers it.
| Where data leaves | What Argent does | Capability |
|---|---|---|
| AI tools nobody approved | Finds them from your web proxy's logs and your managed machines, with the login or device behind each one | Shadow AI Discovery |
| Customer data in a request | Replaces Social Security, card, account and other numbers with a redaction marker before the request leaves | Sensitive Data Redaction |
| Keys and passwords | About 60 credential rules catch cloud, AI-provider and private keys; agents never hold your real provider key | Keyless Agents |
| Data in the model's reply | Redacts or withholds sensitive data the model sends back, and neutralizes image and link tricks that carry data out | Response Screening |
| Hidden instructions in tool results | Marks what an agent's tools bring back as data, and refuses prompts its classifier flags | Prompt Injection Defense |
| An unapproved model or provider | Only the models you approve can answer; a request to any other is refused | Model Control |
| Tools an agent should not touch | An admin decides which MCP servers and tools each agent may use; every other call is refused | Agent Reach Control |
Data protection, credential, tool and model controls each have their own watch switch, so you can see what they would do before they enforce, and every decision is written to the signed record. See Watch Before You Enforce.
The path from first look to enforced
Start by seeing. Nothing changes for staff until you decide.
- See where AI is used. Begin with a copy of your web proxy's access log, in Squid format or NDJSON: nothing is installed and no traffic moves. Then push a setup to managed machines with your device-management tool. See Shadow AI Discovery.
- Route AI traffic through Argent. Any agent or tool with a base-URL or proxy setting can be pointed at the Argent gateway.
- Watch. With data protection in watch mode, Argent works out each redaction, forwards the original, and records what it would have done.
- Enforce. Switch data protection to Enforce. The default policy redacts from the next request, before anyone writes a policy.
- Tune. Add your own patterns and test them on sample text first, fingerprint your real customer lists with Exact Data Match, and set Deny for data that must never leave.
- Close the reply side. Response screening checks what the model sends back. Choose Redact to keep answers flowing, or Block to withhold any reply with a detection.
What it catches, and what happens next
| 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. |
| Restore (opt-in) | The value leaves as a token and is put back into the model's answer, so the person gets a usable reply while the model never saw it. Off until an admin turns it on, and that change is recorded. |
| 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, email, phone, IP address, IBAN, bank account, passport, driver license, medical license and crypto wallet numbers; secrets such as API keys are counted.
- Switch on more: person names, places, dates and web addresses; a pack of about 60 credential rules; US routing numbers and SWIFT codes, checked by checksum.
- Your own formats: an admin writes a custom pattern, tests it on sample text and sees the matches before it goes live.
- Your own records: Exact Data Match fingerprints your real customer or employee lists on your side; only keyed digests are uploaded, never the values.
- Your own engine: if you run your own Microsoft Presidio engine, Argent can send the check to it instead, and refuses the request if that engine fails.
Built to fail closed
A data-loss control that quietly stops checking is worse than none. Argent refuses rather than guesses.
- No check, no request. If the detector is down or a request is too large to scan, the request is refused, not sent.
- No check, no reply. By default, if the reply check cannot run, the reply is withheld.
- Never a silent skip. Turning scanning off is an explicit choice, and it is recorded.
- No empty promises. A policy that lists a protection that can never fire is rejected when it is saved.
- Protected from day one. A company with no saved policy still gets the default protections.
- Checked inside your walls. Detection runs inside your own Argent deployment; the text is not sent to another cloud to be checked.
- Tested on real traffic. No change to the DLP engine is merged until it is replayed against a recorded AI session from Nyedis's own use. It must show zero requests refused because the check could not run, and a 95th-percentile check time under 750 ms, both read from the signed record.
Proof you can hand an auditor
Stopping a leak is half the job. The other half is showing what each AI model was and was not sent.
- A signed line per decision. The record reads like "DLP redacted SSN x1": the kind of data and how many, never the value.
- Two copies, side by side. Replay shows what the agent sent beside what Argent forwarded, with the marker highlighted. The raw copy opens only after a deliberate reveal, and that read is recorded under the reader's name. See Session Replay.
- Captures kept apart. Captured request and reply bodies sit in a separate encrypted store, linked to the signed record only by a digest, kept for as long as you set.
- Counts, not contents. The home screen shows what DLP caught by data type, never the values.
- A package for the auditor. One-Click Audit Evidence builds a signed evidence package they check on their own machine.
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.⁵
- 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."⁶
Where to go next
Go deeper on any part of the path, or start from your use case or industry.
- Related goals: See all AI in your company, Compliance evidence, Audit and accountability, Threat defense, 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.
- Industries: Financial Services, Insurance, Healthcare, Retail, Public Sector and all industries.
- Keep it inside your walls: air-gapped on premises, private cloud or your own cloud VPC. Compare deployment options.
See the whole platform, take the walkthrough, browse all capabilities and integrations, or grab the data sheets.