Assessment
AI Security Review Assistant
Reviews designs, architecture changes and new integrations for security risk, using a threat model.
Assessment
AI Prompt Injection Defense
Attempts to hijack your AI agents or chatbots with hidden or malicious instructions are detected and blocked before they reach the model.
Anyone who can type into your chatbot, or plant text in a page your agent reads, can try to rewrite its instructions. This defense screens messages, retrieved documents and tool outputs before they reach the model. Clear attacks are blocked, borderline ones are flagged for a person, and every attempt is logged. You approve the rules and own the log.
- Injection attempts stopped before the model ever sees them
- Hidden instructions in documents and tool outputs screened too
- Every attempt logged with the rule it matched
$5,400
20–30 working days
Optional maintenance $270/mo
What it is
AI Prompt Injection Defense screens the inputs, retrieved documents and tool outputs reaching your own AI applications and chatbots. It blocks or flags prompt injection attempts. It protects AI apps you already run, whoever built them. It does not cover general application code or infrastructure security.
Why businesses buy it
A chatbot or agent that follows a planted instruction can leak data, insult a customer or call a tool it should never touch. That is a real cost, and it often surfaces only after someone screenshots it. This defense sits in front of the applications you name and checks what reaches the model, including the indirect content most teams forget. The written scope lists which applications are covered and which controls are configured, so you know what is protected before you buy.
Problems this solves
Users try to trick your chatbot into ignoring its instructions
The agent screens inputs reaching your AI applications and chatbots for prompt injection attempts. Suspicious inputs are blocked or flagged before the model acts on them.
Hidden instructions inside documents can steer your AI tools
Retrieved documents are screened as well as user messages. Text planted in a file or web page can be caught before it reaches the model.
Tool outputs feed untrusted text back into your agents
Outputs from tools your AI calls are checked for injection attempts too. An agent is less likely to act on instructions hidden in a tool's response.
You have no record of attempted attacks on your AI
Flagged attempts appear in the admin console for your team to review. You can see what was tried and tune the controls that are in place.
A good fit if
You run a chatbot or AI agent that takes free-text input from customers, website visitors, or the open internet.
Your AI application pulls in documents, web pages or tool outputs that a third party could tamper with to smuggle in instructions.
You want suspicious inputs blocked or flagged before they reach your model, not discovered afterward in a log review.
You already have an AI application live or in testing and specifically want its input and output pipeline defended.
Not the right fit if
We built your AI agent and you only want input and output guardrails added to it. See Prompt Injection Guardrails Add-on, sold alongside our own products.
Your concern is general application code and dependency vulnerabilities, not malicious input to your AI models; see AI Application Security Agent.
You want this bundled with agent guardrails, action limits and centralized monitoring in one build; see AI Agent Security Layer, our broadest option.
Your priority is sensitive data leaving your systems rather than malicious input entering them; see AI Data Leakage Monitor.
You need a one-time assessment of your overall AI security posture rather than a narrow, always-on defense; see AI Security Audit.
How it works
- We connect directly to your AI application's input and output pipeline. This sits in front of the chatbot or agent you name, rather than scanning your wider codebase or infrastructure.
- Detection rules are tuned to your application. We configure pattern and intent checks for known injection techniques, such as instructions embedded in a message asking the model to ignore its prior rules.
- Indirect injection is checked too, not just direct messages. Documents, web pages and tool outputs the AI reads are screened for hidden instructions before that content reaches the model.
- Suspicious inputs are blocked or flagged, based on rules you approve. Clear attempts are stopped before the model sees them; borderline cases are flagged and logged for your review rather than guessed at.
- Every attempt is logged with what triggered it. A record of the input, the rule it matched and the action taken is kept, so patterns of repeated attempts are visible over time.
- Alerts reach Slack so a person can review edge cases. Nothing about this defense claims to stop every possible attempt; it narrows the exposure and gives your team visibility into what is being tried.
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 applications we may put this defense in front of.
- The list of chatbots, agents or AI applications you want in scope.
- Examples of the kind of input your application normally receives, including any known problem cases.
- Read and write access to the application's input and output pipeline, as needed.
- A named IT or security owner who can review flagged attempts 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 detection 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
AI applications, scope and detection rules confirmed.
Weeks 1–2
Input and output screening built and tuned to your application.
Week 3
Pipeline connections and alerting 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 | $5,400 |
| Express | 15–25 working days | $6,200 |
| Rush | 10–20 working days | $7,000 |
What you receive at handover
- Prompt injection screening on inputs, retrieved documents and tool outputs for up to three AI applications
- Up to five controls or policies configured to block or flag attempts
- An admin console for reviewing flagged items
- 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
- Your AI tools can be harder to manipulate.
- Attempts can be seen rather than missed.
- You can point customers to a specific control against prompt injection.
- Attempts blocked or flagged
- Count them each month from the admin console.
- False positives
- Sample flagged items each month and record how many were harmless.
- Injections that got through
- Track any confirmed cases that reached the model, from your incident log.
Before you start: List the AI applications in scope and any known misuse so far. Keep a sample of recent inputs to test against.
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 defense costs roughly $20 to $300 a month in your own accounts, for hosting, logging and the monitoring tools it relies on day to day.
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 AI applications, 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
Monthly maintenance covers monitoring and a monthly allowance of change hours for things such as tuning detection rules or adding a second application. It starts after the 30 days of included support and can be cancelled any month.
$270/mo (5% of the build price a month, includes 2 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 | A guardrail platform charges per request volume, monthly | A security firm quotes after a scoping call | Ours is one fixed price, shown on this page before you buy |
| Scope | A guardrail platform applies one generic rule set across all customers | An agency agrees scope during the engagement | Ours is written on the Scope Card, tuned to your application |
| Delivery | A guardrail platform still needs tuning before it is reliable | A security firm engagement often takes weeks to schedule | Ours has a fixed delivery time shown before you buy, with faster options |
| Coverage | A guardrail platform mainly filters direct user messages | An agency build depends on what was agreed | Ours checks direct input and indirect content like documents and tool outputs |
| Guarantee | A guardrail platform may market a block rate figure | An agency build depends on what was agreed | Ours reduces exposure; no defense blocks every attempt |
Questions about this product
Does AI Prompt Injection Defense stop every injection attempt?
No defense can block every possible attempt, and we do not claim otherwise. It reduces your exposure by screening direct messages and indirect content such as documents and tool outputs against known injection patterns, blocking clear attempts and flagging borderline cases for a person on your team to review.
How is AI Prompt Injection Defense different from AI Application Security Agent?
This defense is narrow and specific to malicious input reaching your AI models, such as messages or documents designed to hijack a chatbot's instructions. AI Application Security Agent covers your broader codebase and dependencies instead. They address different risks, and many teams run both together for full coverage.
What counts as an indirect injection attempt?
It is an instruction hidden inside content your AI application reads rather than typed directly by a user, such as text embedded in a webpage, a document, or the output of a tool the agent calls. This defense screens that content too, not only the direct chat messages a person types.
How much does AI Prompt Injection Defense cost to run each month?
Expect roughly $20 to $300 a month in your own accounts for hosting, logging and the monitoring tools the defense runs on. Most single-application setups sit toward the lower end, and the figure rises gently with traffic. You pay those providers directly, and we add no per-message fee. The one-time build price for each delivery speed is in the price panel on this page.
What happens when an attempt is flagged instead of blocked?
Borderline cases are logged and routed to Slack for a named person on your team to review, rather than silently allowed through or automatically blocked without any oversight. Clear, high-confidence attempts are blocked outright based on the rules you approved during setup. This keeps the defense cautious about ambiguous cases rather than making a risky call on its own.
Can I add a second chatbot or AI application later?
Yes, within the three systems included in scope, adding a second application is part of the build. Extending coverage beyond three AI applications, or substantially different applications added later, comes through a Change Quote, priced and agreed with you before that work starts. This keeps costs predictable as your use of AI applications grows over time.
Who reviews the logged injection attempts?
Your named IT or security owner does, using the log this defense keeps of every attempt, the rule it matched and the action taken. We do not review or act on your logs ourselves; monthly maintenance covers tuning the detection rules if attack patterns change over time.
Stop hidden instructions from hijacking your AI.
Order AI Prompt Injection Defense and it starts screening input and content reaching your AI application, live on a fixed delivery date.
Works well with
Assessment
AI Application Security Agent
Scans application code and dependencies for security issues.
Assessment
AI Security Audit
Reviews your AI applications and agents for common security weaknesses.