AgentCop

Keep your agent under control.

Runtime security for AI agents. Cop sits between your agent and its model, reads every tool call before it runs, and catches prompt injection, tool poisoning and data leaks before they become actions.

Drop to scan
Try
  • Files are read, never run
  • Secrets masked
  • Files not kept after the scan
  • Reads every tool call before your agent runs it
  • Holds risky actions at the desk until you decide
  • Emails you when your agent goes off its beat

Under watch in 2 minutes. Two steps, and you're done.

  1. 1Point your agent at CopSwap two values in its settings: the base URL and the key. Its code stays as it is.
  2. 2Give Cop its real providerPaste your provider's key once. Cop keeps it encrypted and passes each checked call on.

Rather not touch anything? Paste the setup prompt into Claude Code or Codex: it does both steps for you.

What Cop does for each agent.

Agent threat detection

Prompt injection hidden in an email, a web page or a ticket, and tool poisoning in an MCP server's descriptions. Cop names the order, the tool it targets and its MITRE ATLAS technique.

Action control before execution

Cop sees each tool call and its arguments before the agent executes it. In Blocking mode, a high-risk call waits for your Release or Keep it locked.

Secrets and sensitive data

Keys, tokens and private files get caught on their way out: in a tool's arguments, in a URL, in the model's answer. Your agent holds a decoy key, and Cop keeps the real one.

Behaviour drift (UEBA)

Cop learns each agent over its first runs: which tools it calls, in what order, where it sends data. A tool it never used, a new destination or a run unlike the others raises the score.

Any model, one URL

OpenAI, Anthropic, OpenRouter, LiteLLM, Azure OpenAI or any OpenAI-compatible endpoint. Point the agent at Cop's URL with its cop- key, or paste the setup prompt into Claude Code or Codex and it does it for you.

Built on the standards your auditor already reads.

Every finding carries its OWASP or MITRE ATLAS ID, so your CISO, your SOC and your auditor read the same report.

Detection built on
Reviews agents on
  • Microsoft Copilot Studio
  • Microsoft Entra Agent ID
  • Microsoft Agent 365
  • Microsoft Agent Framework
  • Agent Governance Toolkit
  • LangGraph and LangChain
  • Custom Python agents
  • MCP servers
Alerts land in
  • Splunk
  • Microsoft Sentinel
How we test each OWASP agentic risk

This is what an agent under control looks like.

How it works. Three layers score every call. 129 standing orders, each mapped to MITRE ATLAS or OWASP, check what the agent read and what it is about to do. Behaviour analysis compares the call with the agent's habits. The agent's own model lists the ways its mission can go wrong. You pick the mode, agent by agent: Audit alerts you, Blocking holds the call.

  • Calls to reviewed tools pass, logged.
  • A tool whose description changed is blocked before the model sees it.
  • The block raises an alert in your SIEM.
Requestfrom a userguard: identity, tool pins, scopeAgentsvc-support, scopedCRMpinned, readTicketspinned, readDocspinned, readbilling-toolsdescription changedBlocked before the modelhash 602cde ≠ 1933bfSIEM alerttool_description_drift
Model calls on the blocked session: 0Blocked in 4 ms, lab run

Inside the console: one agent, under watch.

Real screens from our lab: a phishing-triage agent with 11 tools, watched by Cop. Click a screen to see it whole.

The agent's page in the Cop console: its state, model and risks, counters for runs, calls, tool calls, risk score and attacks caught, then the background check with a radar and one score per OWASP axis.
Background check. A risk score, then how exposed the agent is, axis by axis.
Its 11 tools ranked by risk, each tagged reads, reads outside text or changes things, then its usual runs and its tool calls day by day.
Its tools, ranked by risk. What each tool can do, and the beat the agent keeps.
The case file: four ways the agent can go wrong, from critical to medium, listed by the agent's own model.
Case file. The ways it can go wrong, listed by the agent's own model.

129 standing orders, scope by scope.

Every rule maps to MITRE ATLAS or OWASP. 41 block: they stop the call before it runs. 76 run in audit: they log it and score the run. 12 are disabled until an agent needs them. The rest stays in the fog: that part is ours.

  • Blocking · 41
  • Audit · 76
  • Disabled · 12
The agent's header in the Cop console: an Audit button next to Release, and a note saying Cop flags and alerts while nothing is stopped, with a link to switch to Blocking.
Audit or Blocking. The owner picks, agent by agent, in one click.
Standing orders for one agent: the rules that fired this week, grouped by scope, each with its mode (audit, blocking or off) and how often it fired. Most rule names are blurred.
Standing orders, for one agent. What fired this week, and whether each rule blocks, audits or is off.
Inputs16 rules
  • E01Override order in what it readaudit
  • E02audit
  • E03audit
  • E04audit
  • E05audit
  • E06audit

+ 10 more in the fog

Actions and effects21 rules
  • A01Acts right after outside textaudit
  • A02audit
  • A03blocking
  • A04blocking
  • A05blocking
  • A06blocking

+ 15 more in the fog

Data13 rules
  • D01New destinationaudit
  • D02blocking
  • D03blocking
  • D04audit
  • D05blocking
  • D06blocking

+ 7 more in the fog

Identity and secrets10 rules
  • I01Secret in an outbound callblocking
  • I02audit
  • I03audit
  • I04blocking
  • I05audit
  • I06blocking

+ 4 more in the fog

Tools and supply chain10 rules
  • O01New toolaudit
  • O02audit
  • O03blocking
  • O04blocking
  • O05audit
  • O06audit

+ 4 more in the fog

Configuration8 rules
  • C01System prompt changed inside a runaudit
  • C02disabled
  • C03audit
  • C04blocking
  • C05audit
  • C06audit

+ 2 more in the fog

Context and memory7 rules
  • M01Memory written from outside textaudit
  • M02blocking
  • M03audit
  • M04audit
  • M05audit
  • M06blocking

+ 1 more in the fog

Resources and cost7 rules
  • R01Tokens per run spikeaudit
  • R02audit
  • R03audit
  • R04audit
  • R05audit
  • R06disabled

+ 1 more in the fog

Between agents6 rules
  • X01Order passed on to othersblocking
  • X02audit
  • X03audit
  • X04audit
  • X05blocking
  • X06blocking
Machine and deployment5 rules
  • P01Sandbox escape attemptblocking
  • P02blocking
  • P03blocking
  • P04blocking
  • P05blocking
Humans7 rules
  • H01Approval asked for an order from outsideaudit
  • H02audit
  • H03audit
  • H04blocking
  • H05blocking
  • H06audit

+ 1 more in the fog

Drift, no attacker (OWASP)7 rules
  • G01Outside its missionaudit
  • G02audit
  • G03audit
  • G04audit
  • G05audit
  • G06audit

+ 1 more in the fog

Behaviour (UEBA)12 rules
  • B01Tool never used beforeaudit
  • B02audit
  • B03audit
  • B04audit
  • B05audit
  • B06disabled

+ 6 more in the fog

Common uses, and what Cop checks.

Phishing-triage agent

It reads reported emails and blocks malicious URLs. A crafted email can make it block your own domain or quarantine a supplier's invoices. Cop checks the target before the block goes out.

Support agent with refunds

It answers customers and issues refunds. One message can ask for a refund to new bank details. Cop holds a payment to a payee it has never seen.

Company assistant on Drive and Slack

It searches shared files and posts in channels. A poisoned document can make it paste private files into a public channel. Cop flags private text heading to a new destination.

Coding agent with production access

It runs commands on your servers. An issue comment can carry curl | sh. Cop stops download-and-run, destructive commands and edits to the agent's own configuration.

Who runs the review.

Every review is run and signed by the same CISSP-certified security engineer, with eight years in critical environments, most of them on the Microsoft security stack. We never name a client or an employer.

  • 147,000users under data loss prevention, export-controlled data included
  • 7,000endpoints under EDR, run from a level-3 SOC
  • CopilotMicrosoft 365 Copilot secured for a regulated asset manager
  • Laba simulated company with Active Directory, SSO, business apps, MCP servers and a SIEM, where each attack scenario is built and rehearsed

Questions teams ask before connecting an agent.

What is AI agent security?

AI agent security is the practice of controlling what an AI agent does with the access it has: the tools it calls, the data it reads and sends, and the instructions it picks up from emails, web pages, tickets or tool descriptions. It goes beyond model safety, because an agent acts: it sends emails, runs commands and issues refunds.

How does Cop protect an AI agent at runtime?

Cop sits between your agent and its model provider, so every request and every tool call the model asks for passes through it. 129 rules mapped to MITRE ATLAS and OWASP check what the agent read and what it is about to do, and behaviour analysis compares each call with the agent's habits. You get an email when the agent goes off its beat; in Blocking mode, a high-risk call waits for your decision.

Do I need to change my agent's code?

No. You change two settings: the base URL your agent sends its model calls to, and its API key, which becomes a cop- key that works only through Cop. You give Cop your real provider key once. If you would rather not touch anything, paste the setup prompt into Claude Code or Codex and it does both steps for you.

Which model providers work with Cop?

OpenAI, Anthropic, OpenRouter, LiteLLM, Azure OpenAI and any OpenAI-compatible endpoint, streaming included. Cop passes each call on in its own format, so any agent that lets you set its base URL and API key can go through it, whatever framework it was built with.

What is MCP tool poisoning?

An MCP server sends each tool's description to the agent, and the model reads it as instructions. A poisoned description hides orders in that text, such as reading a key file and sending it out. Cop reads every tool description in each request, flags hidden instructions, and raises an alert when a description differs from the one it saw before.

How does Cop catch indirect prompt injection?

Indirect prompt injection is an order hidden in content the agent reads: an email, a web page, a ticket, a file. Cop checks each tool result for those orders, then checks the tool call that follows: a command copied from an email, or private data heading to an address the agent never used, is flagged, or held in Blocking mode.

What is the difference between Audit and Blocking mode?

In Audit mode, Cop lets every call through, logs what each rule would have done and emails you about serious findings. In Blocking mode, a call rated high or critical waits until you release it or keep it locked. You choose per agent, in one click, and you can set each rule to Blocking, Audit or Off for that agent.

What does Cop keep, and where?

Cop online keeps each agent's instructions, its tools, its calls and the rules that fired, for 365 days, on our server in the European Union (France). Your provider key is stored encrypted, and the agent only holds a cop- key. Nothing that passes through Cop is used to train a model. The privacy page lists every field.

Which standards does Cop map to?

Each of the 129 rules carries a MITRE ATLAS technique or an OWASP ID. MITRE ATLAS covers attacks on agents (122 techniques in 11 scopes); the OWASP Top 10 for Agentic Applications 2026, ASI01 to ASI10, covers an agent drifting on its own with nobody attacking it. Your CISO, your SOC and your auditor read the same IDs.

Can you connect one of our agents with us?

Yes. Book a guided pilot: we connect one of your agents with you and go through what Cop caught in its first week.