
The answer can carry what the prompt never sent
A careful prompt is only half the trip. An AI model reads account notes, files and tool results, and its answer can repeat a customer's Social Security or card number straight back to the agent.
- OWASP lists sensitive information disclosure as LLM02 in its 2025 Top 10 for LLM applications: LLMs, "especially when embedded in applications, risk exposing sensitive data, proprietary algorithms, or confidential details through their output."¹
- OWASP's LLM05, improper output handling, is "insufficient validation, sanitization, and handling of the outputs generated by large language models before they are passed downstream to other components and systems."²
- In one OWASP attack scenario, the model encodes sensitive data and sends it, "without any output validation or filtering, to an attacker-controlled server."²
Many teams screen only what goes out to the model. Argent screens what comes back too.
What can come back in a reply
A model's reply is not just text on a screen. Agents act on it, save it and pass it on.
- A customer's identity data. The model summarizes an account and repeats the number it was given. Now it sits in a chat window, a ticket or a saved transcript.
- A tool call with data in it. The model asks the agent to run a tool, and the arguments carry a value that should not travel.
- A hidden image or disguised link. The reply holds a markdown image tag or a link that a chat window renders, and the data rides out in the address.
- Content your own rules forbid. A topic or kind of data your compliance team has said stays out of AI answers.
What is at stake: without a check on the way back, sensitive data lands wherever the reply goes, and no one can show what was checked or what was done about it.
What Argent does on the return trip
Argent Strategic Intelligence is the gateway between your agents and the AI models they use. In Enforce, every reply that comes back through Argent is held until it has been checked, then released, changed or withheld; a reply Argent cannot read is recorded as unchecked, never as clean.
In Enforce:
| Check | What Argent does |
|---|---|
| Sensitive data | The same detector and the same company data loss prevention (DLP) policy that screen prompts read what the model wrote: its words, text split across streamed pieces, and the arguments of tool calls it asks for. Technical fields such as ids, model name and usage counts are skipped. |
| Redact | With the Redact setting and the policy action Redact + Audit (both the defaults), the sensitive spans are replaced in place and the rest of the reply is delivered. |
| Block | Every reply with a detection is withheld. The agent gets a refusal (HTTP 403) that names the data types and a trace id, never the data. |
| Image and link tricks | An external markdown image or a disguised link in the model's text is replaced with a fixed marker such as [image blocked by Argent]. The rest of the reply is delivered. |
| Classifier | Optional. An AI classifier scores the reply for unsafe content; a hit is recorded as flagged, and the reply is withheld under Block. |
| Plain-language rules | Your own rules run on the reply too. Block withholds a match, Warn raises an alert, Shadow only records. |
Response inspection is on by default. A company can turn it off in Add-ons, and that choice is recorded on each request.
How it works, step by step
A bank's relationship-manager assistant asks a model to summarize a customer's account before a banker's call.
- The request goes out through Argent. What the assistant sends is screened by its own control, Sensitive Data Redaction.
- The reply is held. In Enforce, Argent keeps the whole answer until it is complete, and no part of it reaches the assistant before the check. Streamed pieces are joined, so a number split across two pieces is still found.
- The model's words are checked. The reply text and any tool-call arguments go through the company's DLP policy.
- The company's setting is applied. Under Redact with Redact + Audit, the number is swapped for
<REDACTED>and the summary arrives with everything else intact. - Or the reply is withheld. Under Block, the assistant receives
Argent DLP blocked the response: US_SSN detected (trace ...). The data types are named; the data is not. - A signed record is written. One record per inspected reply says what was found and what was done, as categories and counts, never the value.
- No check, no release. In Enforce, if the check cannot run, the reply is withheld with
Argent withheld the response under egress policy (trace ...).
What the operator sees
The controls sit on the DLP page of the Argent console, in the console's Egress, Response lane.
- Response inspection panel. A status pill that reads RESPONSE INSPECT ON, and the note that the DLP policy "also applies to responses before they leave the gate."
- Egress leak handling. Inherit, Redact or Block, with a status line that shows the effective setting and the platform default. It is set per company, and only a SuperAdmin or that company's TenantAdmin can change it.
- Platform default. The Platform Defaults page sets the response leak mode for companies that inherit it.
- Replay. Open a request whose reply was checked and Replay says "DLP scanned the provider's reply on the way back to the agent," and names egress inspection when a reply was withheld. See Session Replay.
- The signed record. Decision, data categories and counts for each reply. Hand it to an auditor with One-Click Audit Evidence.
Redact or Block: a deliberate choice
Redact keeps the answer flowing. Block holds it back. Your company decides, and the record shows which one ran.
| Setting (in Enforce) | What the agent receives | What the record says |
|---|---|---|
| Redact, policy Redact + Audit (defaults) | The reply with the sensitive spans replaced | Redacted |
| Redact, policy Audit Only or Deny | The original reply, unchanged | Detected |
| Block, any policy action | A 403 naming the data types and a trace id | Blocked |
| Watch mode | The original reply, unchanged | What Argent would have done |
| Check cannot run (Enforce default) | A 403 saying the reply was withheld | Withheld |
The console spells out the trade before you pick Block: "Every reply with a detection is refused with a 403, including false positives. Agents that relied on redacted replies will see errors instead."
Who it serves
Any team whose AI agents read customer, patient or account data and pass the answers on.
Response screening is one half of threat defense; the other half is Prompt Injection Defense, which checks what goes in.
- Goals: Data loss prevention, Audit and accountability, Compliance evidence, See all AI in your company 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, Healthcare, Insurance, Retail, Public Sector and all industries.
Where the pressure is written down.
- New York DFS, 23 NYCRR 500.17 (covered entities): notice of a cybersecurity incident "as promptly as possible but in no event later than 72 hours after determining that a cybersecurity incident has occurred," including at a third-party service provider.³
- HIPAA Security Rule, 45 CFR 164.312(b) (covered entities and business associates): "Implement hardware, software, and/or procedural mechanisms that record and examine activity in information systems that contain or use electronic protected health information."⁴
How a team starts
Watch first. Nothing changes for staff until you decide.
- Watch. New companies start in Watch. Every reply is delivered unchanged, and the record shows what Argent would have done. See Watch Before You Enforce.
- Enforce. Switch the DLP area to Enforce. Under the default Redact + Audit policy, Redact starts blanking sensitive spans in replies.
- Choose. Keep Redact, or set Egress leak handling to Block for a company where no flagged reply may pass.
- Review. Read reply-by-reply decisions in Replay and the signed record.
Argent runs inside your own environment, including air-gapped on premises; compare deployment options. See the whole platform, take the walkthrough, or browse all capabilities.