Controls & Monitoring
AI Agent Risk Monitor
Monitors agent actions for risky or unexpected behavior.
Controls & Monitoring
AI Agent Security Layer
Every AI agent you run sits behind one policy layer that limits its tools, screens its input, and logs everything it does.
Running several AI agents means several sets of rules, logs and blind spots. This layer puts one policy wrapper around every agent in scope. It limits which tools each agent may call, screens incoming content for injection attempts, routes sensitive actions for approval and keeps one central log. Your team sets the rules and keeps every record.
- One set of guardrails applied across every agent in scope
- Input screening and tool limits enforced before actions run
- A single central log of inputs, tool calls and outcomes
$8,900
20–30 working days
Optional maintenance $445/mo
What it is
AI Agent Security Layer is one layer wrapping every AI agent you run, combining tool and action guardrails, input screening against prompt injection, and centralized logging into a single build. It is the broadest option in this cluster, in place of running each control separately.
Why businesses buy it
Each new agent tends to arrive with its own guardrails, built in a hurry, and its own log in a different place. That works for one agent. With three, nobody can say for sure what any of them is allowed to do. This layer replaces that patchwork with one consistent set of controls and one record. It reduces risk rather than removing it, and the written scope lists exactly which agents, controls and systems are covered before you commit.
Problems this solves
AI agents can take actions nobody intended them to take
Tool and action guardrails set what each agent may do. Actions outside those limits are stopped, so an agent cannot wander beyond its intended job.
Each agent has its own patchwork of security controls
One layer wraps the agents in scope, combining guardrails, input screening and centralized logging. You manage a single build instead of separate controls for each agent.
Prompt injection reaches agents through inputs and documents
Input screening checks what reaches each agent for prompt injection attempts. Suspicious inputs are caught at the layer before the agent can act on them.
There is no single log of what your agents did
Centralized logging records agent activity in one place. When something goes wrong, your team can trace what happened without collecting logs from each separate tool.
A good fit if
You run more than one AI agent and want them all governed by the same set of rules, not separate configurations.
Your agents can call tools or take actions, such as sending messages or writing to a system, and you want limits enforced.
You want prompt injection screening, action guardrails and centralized logging together, rather than buying and integrating each piece separately.
You have outgrown a single narrow control and need something that covers your whole AI agent footprint as it grows.
Not the right fit if
You only need input screening for one chatbot, not a layer across multiple agents; see AI Prompt Injection Defense, which is narrower and cheaper.
You only need to control what actions your agents can take, not the full combination; see AI Agent Permission System.
You want a one-time point-in-time report before deciding what to build; see AI Security Audit, as a starting assessment.
You need written policy documents rather than a technical enforcement layer; see AI Security Policy System, for governance and documentation.
You only want alerts when agent behavior drifts, for a person to review, without enforcement. See AI Agent Risk Monitor.
How it works
- We connect to every AI agent you name, up to the systems in scope. Each one is registered with the layer rather than treated as a one-off integration, so rules apply consistently.
- Tool and action guardrails are set per agent. For each agent, we define which tools it may call and which actions require no approval, a logged approval, or are blocked outright.
- Input screening runs in front of every agent. Messages and content reaching each agent are checked against known injection patterns before the model sees them, blocking clear attempts and flagging borderline cases.
- Guardrails are enforced at the point of action, not after the fact. When an agent attempts something outside its permitted rules, the layer stops it or routes it for approval before it executes.
- Every input, tool call and outcome is logged centrally. Instead of separate logs per agent, activity across your whole AI agent footprint lands in one place you can review.
- Alerts for anything unusual reach Slack or Jira. Repeated blocked attempts, a new pattern of tool use, or an agent approaching its limits are surfaced to a named person, not left buried in the log.
What you provide within 72 hours
Your portal lists these the moment you order. Delivery starts the working day after everything is in and your access is checked.
- Written authorization confirming which AI agents and systems we may wrap with this layer.
- The list of AI agents, the tools each one can call, and the systems you want in scope.
- Your approval rules for actions that should require a person before they execute.
- Read and write access to the agents' tool-calling and messaging pipeline, as needed.
- A named IT or security owner who can review flagged activity after launch.
- Admin access and API keys for every connected system, shared by adding our team as a user or creating a key you control, never by email.
- One decision-maker who can approve the guardrail and screening rules and sign off the launch.
The 72-hour rule. Send everything on your checklist, including access, within 72 hours of ordering so your build starts on time.
Timeline
Start
Agents, tools and guardrail rules confirmed.
Weeks 1–2
Guardrails, screening and logging built for each agent.
Week 3
Agent, tool and alerting connections completed.
Weeks 3–4
Documented test scenarios run and results sent to you.
Week 4
Your review and two rounds of changes to the rules.
Go-live
Live, recorded handoff session, 30 days of support begin.
| Option | Delivery time | Price |
|---|---|---|
| Standard | 20–30 working days | $8,900 |
| Express | 15–25 working days | $10,000 |
| Rush | 10–20 working days | $11,150 |
What you receive at handover
- A security layer around up to three AI agents or applications
- Tool and action guardrails, input screening against prompt injection, and centralized logging
- Up to five controls, policies or monitors configured from your policies
- An admin console or written report, plus a control test log from testing
- An admin guide, a recorded handoff session and 30 days of post-launch support
Expected benefits and how to measure them
- Agents can be kept within the limits you set.
- Agent security can be managed in one place.
- Incidents can be traced through one log.
- Blocked actions and inputs
- Count blocked or flagged events per month from the central log.
- Agent incidents
- Track incidents involving agents in your incident log, before and after launch.
- Time to investigate
- Log how long it takes to trace what an agent did during an incident.
Before you start: List the agents in scope, what tools each can use and how they are monitored today. Record any past incidents.
These are things to track, not promised results. Outcomes depend on your data, your team and your market.
Running costs
The build price is one-time. Running the layer costs roughly $20 to $300 a month in your own accounts, for hosting, logging and the monitoring tools it relies on day to day, scaling with how many agents and how much activity you run through it.
AI model usage during the build and testing is billed the same way, directly to your accounts, so you keep ownership of every provider relationship involved.
Ask about running costsCustom requirements
If you need something outside the Scope Card, such as coverage across more than three systems or a larger set of agents, add a note at checkout. You get a written Change Quote, and that work only starts once you accept it. Larger or bespoke projects go through the Custom AI request form instead.
Request a custom versionKeep it running
We recommend monthly maintenance, which covers monitoring and a monthly allowance of change hours for things such as adding a new agent or adjusting guardrail rules. It starts after the 30 days of included support and can be cancelled any month.
$445/mo (5% of the build price a month, includes 4 change hours)
Optional. Starts after the post-launch support period. Cancel any month.
How maintenance worksHow this compares
| Do it yourself with a SaaS tool | Hire a custom agency | Upmark Labs | |
|---|---|---|---|
| Price | Buying separate point tools for guardrails, screening and logging adds up monthly | A security firm quotes after a scoping call | Ours is one fixed price, shown on this page before you buy |
| Scope | Separate tools each need their own integration and configuration | An agency agrees scope during the engagement | Ours is one Scope Card covering guardrails, screening and logging |
| Delivery | Stitching separate tools together typically takes weeks longer to get right | A security firm engagement often takes weeks to schedule | Ours has a fixed delivery time shown before you buy, with faster options |
| Consistency | Separate tools can drift out of sync across different agents | An agency build depends on what was agreed | Ours applies the same rules across every agent in scope |
| Guarantee | A guardrail platform may market a block rate figure | An agency build depends on what was agreed | Ours reduces exposure; no layer eliminates all risk |
Questions about this product
What does AI Agent Security Layer actually combine?
It combines three things in one build: tool and action guardrails that limit what each agent can do, input screening that checks for prompt injection before the model sees a message, and centralized logging of activity across every agent in scope. Buying it as one layer keeps the rules consistent.
How is AI Agent Security Layer different from AI Agent Permission System?
AI Agent Permission System focuses specifically on what actions each agent is allowed to take, action by action. AI Agent Security Layer includes that plus input screening against prompt injection and centralized logging across your whole agent footprint, making it the broader and more expensive of the two options.
Do I need this if I already have AI Prompt Injection Defense?
If you only run one agent and just need input screening, the standalone Prompt Injection Defense product may be enough and costs less. This layer makes more sense once you are running several agents and want guardrails, screening and logging applied consistently across all of them from a single build.
How much does AI Agent Security Layer cost to run each month?
Expect roughly $20 to $300 a month in your own accounts for hosting, logging and monitoring, scaling with how many agents and how much activity runs through the layer. That range covers most multi-agent deployments comfortably as usage and agent count grow.
Does the layer ever act on its own without a person?
It enforces the rules you set, such as blocking an action outside an agent's permitted scope, but anything ambiguous or high-stakes routes to a named approver rather than proceeding automatically. We do not give the layer authority to make judgment calls that should always sit with a person.
Can I add a new agent to the layer later?
Yes, within the three systems included in scope, adding an agent is straightforward and fits inside the monthly change hours that come with maintenance. Extending coverage beyond three systems, or a much larger set of agents, comes through a Change Quote, priced and agreed with you before that work starts. This keeps the cost of expanding coverage predictable as your agent footprint grows over time.
Who owns the guardrail rules and logs once it is live?
You do. The guardrail configuration, the screening rules and the centralized activity log all live in your own connected accounts, not a separate platform we control, so you retain full access and can adjust the setup at any time you choose. We keep no separate copy that would lock you into working with us going forward.
Put one policy layer around every agent you run.
Order AI Agent Security Layer and it starts governing your AI agents' tools, input and logs together, live on a fixed delivery date.
Works well with
Controls & Monitoring
AI Security Policy System
Maintains security policies and checks systems against them.
Controls & Monitoring
AI Security Monitoring
Centralizes security events from AI systems with alerting.