Your goal: defend your AI against attack

ARGENT FORTHREAT DEFENSE

Book a demo

Master Control System · by Nyedis

01 / AI THREAT DEFENSE

Your goal: keep attackers from steering your AI

If your goal is to defend your AI against attack, start where the attacks travel: in the words your agents read, and in the answers that come back. An AI agent fetches web pages, opens documents and calls tools, and whatever comes back goes straight into the AI model's next prompt.

  • OWASP lists prompt injection as LLM01, the first risk in its 2025 Top 10 for LLM applications.¹
  • OWASP's Top 10 for Agentic Applications 2026 puts agent goal hijack first: "Hidden prompts turned copilots into silent exfiltration engines."²
  • AI models "do not enforce a security boundary between instructions and data inside a prompt," the UK National Cyber Security Centre (NCSC) wrote in December 2025.³
  • IBM found 92% of organizations with an AI-related breach lacked proper AI access controls.⁴

Argent gets you there from one checkpoint: it marks what goes in, checks what comes back, watches every agent, and lets a person contain one on its next request.

02 / AI THREAT DEFENSE

Where AI attacks come in, and what answers each

Argent Strategic Intelligence is the gateway between your agents and the AI models and tools they use. It covers every agent that works through Argent.

The threatWhat Argent doesCapability
Planted orders in a web page, file or tool replyWraps the text in a marker naming its source and tells the model it is data. Optional screening refuses a flagged tool result.Prompt Injection Defense
A prompt that tries to override the rulesThe Classifier, or your own guardrail, scores it and refuses a hit before anything reaches the model.Prompt Injection Defense
A reply that carries data outIn Enforce, holds the reply until it is checked, then redacts the sensitive spans or withholds it. A hidden image or disguised link is replaced with a fixed marker.Response Screening
A poisoned or changed MCP toolEach tool waits for a person to approve it, and its definition is scanned for planted instructions. A tool changed after approval waits again.Agent Permissions
A hostile or changed skillSkills an agent loads are recorded, scanned and compared with the approved version. In enforce, only reviewed versions run.Skills, below
An agent acting out of characterEach agent gets its own baseline. Drift is recorded and shown, and a person, or a policy you chose, holds the agent.Agent Behavior Watch
Repeated attemptsFive injection refusals by one agent inside ten minutes raise a detection and that agent's threat score.Prompt Injection Defense
03 / AI THREAT DEFENSE

Going in: label it, or refuse it

Two doors bring planted orders in: what the agent reads, and what a person types.

  • Untrusted Fencing. Tool results, MCP tool replies and agent-to-agent replies are wrapped in a random marker with a source tag, and the model is told the text is data, not orders. A fake end marker hidden in the text is defaced first. Switched on for every company by default. Once the Untrusted Fencing area is in Enforce, a request that cannot be fenced safely is refused rather than sent on unfenced.
  • Tool-result screening. An opt-in setting that rates tool results for planted orders before they reach the model, and refuses a flagged one with classifier.tool_result_injection.
  • Classifier. A separate AI model with one job, switched on where your deployment runs a scorer. It scores the newest prompt from 0 to 1 for prompt injection, jailbreak and other harms. At or past the line your team sets (0.7 by default) the request is refused with classifier.prompt_injection, and nothing reaches the model.
  • Your own guardrail. Hand the decision to your AWS Bedrock Guardrails or LLM Guard. If the vendor does not answer, the request is refused; it never quietly falls back.
  • Plain-language rules. Your team writes rules about a topic or a kind of data and sets each to Shadow, Warn or Block.

The film above walks one hidden line on a borrower's web page through every step. Read the whole layer on Prompt Injection Defense.

04 / AI THREAT DEFENSE

Coming back: check the reply before the agent acts on it

A model's reply is not just text on a screen. Agents act on it, save it and pass it on. OWASP lists sensitive information disclosure as LLM02: models embedded in applications "risk exposing sensitive data, proprietary algorithms, or confidential details through their output."⁵

  • Held until checked. Once a reply check is in Enforce, every reply that comes back through Argent is held until it is complete and checked; in Watch, replies flow through and a copy is checked. Streamed pieces are joined, so a number split across two pieces is still found.
  • Same policy both ways. The detector and company DLP policy that screen prompts read the model's words and the arguments of any tool call it asks for.
  • Redact or Block, your choice. Under Redact with the policy action Redact + Audit (both the defaults), sensitive spans are replaced and the rest of the reply is delivered. Under Block, every reply with a detection is withheld, and the agent gets a refusal that names the kinds of data found and a reference number, 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].
  • Your rules and the Classifier. Plain-language rules run on the reply too, and the optional Classifier flags a reply and, under Block, withholds it.
  • No check, no release. In Enforce, if the check cannot run, the reply is withheld by default.

See Response Screening and its outbound twin, Sensitive Data Redaction.

05 / AI THREAT DEFENSE

Around the agent: tools, skills, behavior and containment

An attack does not have to arrive in a prompt. It can sit in a tool the agent trusts, a skill it loads, or a change in how it behaves.

ControlWhat Argent does
MCP toolsArgent reads each MCP server's tool list and records a fingerprint of every tool's definition. A new tool, or one whose definition changed after approval, waits for a decision. A scanner flags planted instructions, suspicious file paths and look-alike characters in a definition, and records the finding types, never the matched text. Servers your agents use that nobody registered are listed, so each can be approved or denied.
SkillsEach skill an agent loads is recorded by version and scanned for shell, network, credential-path, injection and hidden-payload patterns. A new version is compared with the approved one. Skill mode runs Off, Watch or Enforce, and every approval is signed.
BehaviorArgent learns ten behaviors per agent, such as its tools, request rate and source networks, and records drift when an agent stops acting like itself. See Agent Behavior Watch.
Detections and threat scoreRepeated refusals, new tools, spend spikes and other patterns become detections with counts and record ids only. Each agent carries a threat score with the list of what added up to it, and the threat map has a prompt injection lens.
ContainmentAdmins can Pause, Quarantine (every request refused until released) or Kill (the key stops working for good) an agent from its page. Owners can quarantine their own agents.
06 / AI THREAT DEFENSE

The path: watch, enforce, then contain

Watch first. Nothing changes for your people until you decide.

  1. Fencing is switched on from day one, in Watch. New companies start every area in Watch: the record shows which tool results Argent would have fenced, and agents still get the original text until the Untrusted Fencing area moves to Enforce.
  2. Watch the rest. Set the Classifier to Fire + Audit and the Observe tab shows what it would have flagged on live traffic, with nothing blocked. Plain-language rules start in Shadow. New companies start in Watch, so replies are delivered unchanged and the record shows what Argent would have done. See Watch Before You Enforce.
  3. Approve tools and skills. Review the pending MCP tools with their scan results, approve the ones each agent needs, and grant them per agent. See Agent Permissions.
  4. Enforce, one control at a time. Switch the Classifier to Sync Block, move rules from Shadow to Block, turn on tool-result screening where you want a hard stop, and set replies to Redact or Block. Each one starts refusing when its area (Untrusted Fencing, Classifier, MCP Gate or DLP) moves from Watch to Enforce. That move takes a fresh passkey check, and the other areas keep watching.
  5. Review and contain. Detections and threat scores show which agents keep drawing refusals, with the signed record behind each one. Quarantine or kill an agent and its next request is refused.
07 / AI THREAT DEFENSE

Proof points: every decision on the record

Defending is half the job. Showing what was defended, months later, is the other half.

  • Codes and counts, never the text. A refused prompt goes on the signed record as its refusal code beside its score, such as classifier.prompt_injection, never the prompt itself.
  • Fencing is recorded. The record notes that fencing applied and how many pieces were fenced.
  • One record per inspected reply. What was found and what was done, as categories and counts, never the value.
  • Detections without re-exposure. Detection rows hold counts and record ids only, and a high-severity one opens a case.
  • Replay. Open any request and Replay shows what each check did, including when a reply was withheld on the way back. See Session Replay.

Built the way the NCSC advises. The NCSC warns that "the best we can hope for is reducing the likelihood or impact of attacks."³

NCSC adviceHow Argent answers it
Mark data apart from instructionsUntrusted Fencing labels every tool result with its source, inside a random marker an attacker cannot fake.
Use controls outside the modelThe gateway refuses by code and threshold before the model is called. In Enforce, if a check cannot run in time, the request is refused, not passed.
Log enough to spot abuseEvery decision goes on a signed, tamper-evident record, and repeated refusals raise a detection.
Limit what an agent can reachEach tool is approved one by one, so a fooled agent has less within reach.

Hand it to an auditor with One-Click Audit Evidence.

08 / AI THREAT DEFENSE

Where to go next

Argent Strategic Intelligence

Book a demo

See Argent Strategic Intelligence stop a real AI agent, live, with the engineers who built it.