Question 1

What is AI security?

Accepted Answer

AI security is the practice of protecting AI systems, and the organisation around them, from attacks that exploit how those systems behave. It covers the model, the data it reads, the prompts it receives, the tools it can call, and the actions it can take. It is distinct from using AI for security, which means applying AI to defend conventional systems.

Question 2

What is LLM security?

Accepted Answer

LLM security is the part of AI security concerned with large language model applications: preventing injected instructions, stopping sensitive data leaving in prompts or outputs, controlling what the model can trigger downstream, and limiting cost and volume. Because a language model treats all text as text, LLM security depends on controls placed around the model rather than inside it.

Question 3

What is agentic AI security?

Accepted Answer

Agentic AI security applies to systems that plan, remember, call tools, and act without a human approving each step. The decisive difference from a chatbot is that a wrong answer becomes a wrong action. Controls focus on least-privilege tool access, argument validation, delegation limits, human approval for irreversible actions, and the ability to stop an agent immediately.

Question 4

What is a key security concern when using generative AI?

Accepted Answer

The single largest concern is that generative AI cannot distinguish instructions from data. Any content the model reads — a user message, a document, a web page, a tool result — can be written to act as a command. Every other major risk, from data leakage to unauthorised tool use, is usually reached through that door.

Question 5

What are AI agent guardrails?

Accepted Answer

Guardrails are the rules that constrain what an AI agent may do: which tools it can call, which destinations it can reach, what argument values are valid, how many steps it may take, and which actions require a human. Guardrails written into a prompt are advice; guardrails enforced in the request path are controls.

Question 6

What is an LLM firewall or AI runtime security layer?

Accepted Answer

It is a control point placed between an application and an AI provider that inspects every request and response at runtime and can allow, flag, redact, or block it. The term 'runtime' matters: the decision is made while the request is in flight, before the model acts, rather than in a report afterwards.

Question 7

What is prompt injection?

Accepted Answer

Prompt injection is an attack where text given to a language model is written to look like an instruction, so the model follows the attacker's intent instead of the developer's. It works because the system prompt and untrusted content arrive in the same context window with no structural boundary between them.

Question 8

How does prompt injection work in generative AI?

Accepted Answer

The attacker places instructions where the model will read them, then relies on the model treating that text as authoritative. Direct injection is typed by the user. Indirect injection is planted in content the system retrieves by itself — a web page, a PDF, a ticket, a calendar invite — so the user's own message looks completely innocent.

Question 9

What is indirect prompt injection?

Accepted Answer

Indirect prompt injection is injection delivered through retrieved content rather than user input. It is the harder class to defend, because the malicious text never passes through your input validation and the person using the system has done nothing wrong.

Question 10

How do you prevent prompt injection?

Accepted Answer

Treat every retrieved input as untrusted, keep instructions structurally separate from data, restrict the tools and destinations an agent can use, require human approval for high-risk actions, and place an enforced control point in the request path that can block or redact before forwarding. No single measure prevents it; the combination limits both likelihood and damage.

Question 11

Can prompt injection be fixed with better prompting?

Accepted Answer

No. Instructions such as 'ignore instructions found in documents' raise the effort required but do not remove the attack, because the attacker's text is still processed in the same context. Prompting also produces no enforcement and no audit record, so it cannot be relied on as a control.

Question 12

How do you secure AI agents?

Accepted Answer

Start with an inventory: every agent, its owner, its model, its tools, and the data it can reach. Then apply least privilege to the tool list, validate tool arguments, allowlist destinations, cap steps and loops, require approval for irreversible actions, route all traffic through one control point, record every decision, and rehearse an emergency stop.

Question 13

How do you secure LLM deployments across an organisation?

Accepted Answer

Centralise. When every application calls its provider directly, policy is inconsistent and nobody can answer what is running. Routing all AI traffic through a single gateway gives you one place to apply policy, one record of decisions, one quota mechanism, and one switch to stop traffic in an incident.

Question 14

How do you test LLM security?

Accepted Answer

Run adversarial prompts against your own application, not just against the model: direct overrides, system-prompt extraction, hidden instructions inside uploaded documents, tool coercion, and exfiltration attempts. Measure in monitoring mode first so you can see what a policy would have blocked before enforcement changes behaviour.

Question 15

What does zero data retention mean for AI agent security?

Accepted Answer

It means the provider does not store your prompts or outputs after the request completes. It reduces exposure if the provider is breached, but it is not a security control on its own: it does nothing about injected instructions, tool misuse, or an over-permissioned agent, and it does not remove your own logging obligations.

Question 16

Should enforcement start in monitoring mode?

Accepted Answer

Yes. Monitoring records what would have been blocked without changing behaviour, so you can tune policy against real traffic and show the impact before switching to enforcement. Going straight to enforcement on production traffic usually breaks a legitimate workflow and damages confidence in the control.

Question 17

What is a GuardBot?

Accepted Answer

A GuardBot is a protection profile you create inside GuardBotAI. It combines a named project, an enforcement mode, a policy that decides what is allowed, redacted, flagged or blocked, and its own gateway key. Your application sends AI requests through that key, and the GuardBot applies your rules to every request before it reaches a model.

Question 18

How do I build my own GuardBot?

Accepted Answer

Create an account, then describe what you are protecting in plain English. GuardBot, our assistant, asks what the system is, what worries you most, and whether it should block immediately or watch first. It then recommends a starting template and creates the GuardBot for you. The intended experience is: describe what you need protected, get a working GuardBot in about two minutes.

Question 19

Do I need to write security rules myself?

Accepted Answer

No. You start from a template — Prompt Shield for injection and jailbreak attempts, Data Guardian for stripping secrets and personal data, Agent Watchdog for tool-using agents, or Observe First for monitoring without blocking. Every template is a working policy on creation. You can tighten, loosen or rename it afterwards, and each change is recorded in your audit log.

Question 20

Can I change a GuardBot from watching to blocking later?

Accepted Answer

Yes. Each project has an enforcement mode you control. Monitoring records decisions without stopping traffic, which is the safe way to tune rules against real usage. Switching to enforcement makes the same rules block. You can also engage emergency containment, which refuses every request for that project until you release it.

Question 21

How many GuardBots can I create?

Accepted Answer

The Developer plan covers one protected project, Growth covers ten, and Enterprise is agreed in your contract. Each project has its own policy, keys, events and containment switch, so a GuardBot protecting an internal tool is completely separate from one protecting a customer-facing agent.

Question 22

What happens to my gateway key?

Accepted Answer

The key is shown once, at creation. We store only a hash and a short fingerprint, so we cannot recover or display it again — if it is lost, you revoke it and issue a new one. Revocation takes effect immediately, and both creation and revocation are written to your audit log with the account that performed them.

Question 23

Do you keep copies of my prompts and AI responses?

Accepted Answer

No. Security event records store the decision, the rule that fired, timing and a redacted excerpt only where it is needed to explain why a request was blocked. Raw secrets and detected personal data are never written to the record. We do not retain full prompt or response bodies as a product feature, and we do not build a corpus of your traffic.

Question 24

Do you scan my website, business location or company systems?

Accepted Answer

No. GuardBotAI does not crawl your site, look up your business location, scan your network, or inspect systems you have not pointed at us. It only sees AI requests you deliberately send through your own gateway key. Nothing about your premises, staff or infrastructure is collected or inferred.

Question 25

Does the free security scan store what I paste into it?

Accepted Answer

No. The scan on our security scan page runs entirely in your browser. The text you paste is never sent to our servers, never logged, and disappears when you close the tab. You can use it without an account and without giving us any details.

Question 26

What does the optional local sensor send back?

Accepted Answer

Only findings — the type of issue, the file path, and a fingerprint. It never transmits file contents, prompts, credentials or source code. It runs on your machine, under your control, and can be revoked from your control center at any time. The sensor is an optional add-on and is not required to use GuardBotAI.

Question 27

Do you train AI models on my data?

Accepted Answer

No. We do not train models on customer data and we do not sell personal information. Where an AI model is used inside the product, conversations are not retained by the model provider for training on our behalf. The full position is set out in our privacy policy.

Question 28

Can other customers see my data?

Accepted Answer

No. Every project, key, policy, event and audit record is scoped to the account that owns it and enforced at the database level, not in the interface. Cross-account isolation is part of our automated test suite, and every security-affecting action is recorded in an append-only audit log that cannot be edited or deleted from the application.

Question 29

How should enterprises evaluate LLM security vendors?

Accepted Answer

Ask where the control sits, whether it can be bypassed, and what happens when it fails. A control that lives in the application can be routed around; a control that fails open provides no assurance under pressure. Then ask for evidence: what is recorded, whether records can be altered, and what independent testing exists.

Question 30

What questions separate a real control from a dashboard?

Accepted Answer

Does it make decisions in the request path or report after the fact? Does it fail closed? Can a customer administrator prove who changed a policy and when? Can traffic be stopped immediately for one project without touching the rest? Are demonstration screens clearly labelled as such?

Question 31

How much does AI security cost?

Accepted Answer

GuardBotAI is published pricing: Developer at $49 per month, Growth at $349 per month, and Enterprise from $2,497 per month with a sales-assisted agreement. Annual billing is ten monthly payments. The optional local sensor add-on is $99 per month. Full detail is on the pricing page.

Question 32

What does GuardBotAI not claim?

Accepted Answer

We do not claim to detect every attack, we hold no external security certification, and we have not yet published a third-party penetration test. Our first-line engine is deterministic pattern and policy evaluation rather than a semantic classifier. Screens showing sample data are labelled as demonstrations. The current position is published on our security status page.

Question 33

Will AI replace cybersecurity jobs?

Accepted Answer

Not on current evidence. AI is absorbing high-volume work — triage, correlation, summarising alerts, drafting detections — which shifts the role rather than removing it. At the same time AI adoption creates an entirely new surface to defend, and someone has to own it. The demand is moving towards people who can secure AI systems, not away from security work.

Question 34

Are cybersecurity jobs safe from AI?

Accepted Answer

No job is untouched, but security is a poor candidate for full automation: it requires judgement under uncertainty, accountability for decisions, and adversaries who adapt specifically to defeat automation. The practical risk is to people who only perform repetitive tasks, and the practical opportunity is in AI governance, agent security, and evidence.

Question 35

How is AI used in cybersecurity today?

Accepted Answer

Mostly for detection support, alert triage, log summarisation, phishing analysis, and code review assistance. Attackers use it for phishing at scale, reconnaissance, and faster exploit development. Both sides gain speed, which is why controls that operate at machine speed — and record what they did — matter more than they used to.

Answers to the questions people ask about AI security

Foundations