Inventory and roles
Which agents you run, what they read, what they may do, who the vendor is. Your role under the AI Act (provider or deployer).
AI agent risk audit
Where your AI agent can fail, and how much it would hurt.
We test for prompt injection, personal data leaks, unauthorised actions and out-of-scope answers. We document the results against the EU AI Act and the AIUC-1 standard and prepare material for your management, data protection officer, auditor and insurer.
This is not an “AI audit” that looks for places to use AI. We test the security of the agent you already have, whoever built it.
Risk map
Describe your agent and the map rearranges itself.
Risk grows with what the agent reads, what it may do and who it answers. Human approval and logs bring it down. Click a dot for details of the test.
For this profile, 1 of 9 risks come out high. Highest: personal data leak.
Risk 1 · Low · Security
Direct prompt injection
How we test
Dozens of persuasion variants: role-play, “ignore previous instructions”, split instructions, another language.
What helps
Separated system instructions, an allow-list of actions in MCP, human approval.
Six areas of the AIUC-1 standard · we do not issue certificates
Data & privacySecuritySafe outputsReliabilityAccountabilitySociety
Risk ranking for this profile
- 3Personal data leakHigh
- 7No audit trail: who, what, whenMedium
- 5Hallucinated answersMedium
- 6Out-of-scope answersLow
- 9Unfair or inappropriate treatmentLow
- 1Direct prompt injectionLow
- 8Outage or silent model changeLow
- 2Indirect prompt injectionLow
- 4Unauthorised or irreversible actionLow
Illustrative map, not an audit resultscore = likelihood × impact
All 9 risks: how we test them and what helps
1. Direct prompt injection
How we test: Dozens of persuasion variants: role-play, “ignore previous instructions”, split instructions, another language.
What helps: Separated system instructions, an allow-list of actions in MCP, human approval.
AIUC-1 area: Security
2. Indirect prompt injection
How we test: An instruction hidden in an e-mail, a PDF or a web page the agent reads. We test on your own document types.
What helps: Untrusted content is labelled; no actions triggered by document content without a person.
AIUC-1 area: Security
3. Personal data leak
How we test: Attempts to extract other people’s data, internal documents or the system instructions.
What helps: Data minimisation and pseudonymisation, roles, output filtering, a DPIA.
AIUC-1 area: Data & privacy
4. Unauthorised or irreversible action
How we test: We try to get the agent to send, delete or change something outside its scope.
What helps: Read-only by default, an allow-list of actions, nothing irreversible without a person.
AIUC-1 area: Accountability
5. Hallucinated answers
How we test: A set of control questions with known answers; we measure the share of errors and invented sources.
What helps: Answers only with sources, clean data, “I don’t know” as an allowed answer.
AIUC-1 area: Reliability
6. Out-of-scope answers
How we test: Legal advice, promises on behalf of the company or authority, inappropriate content, topics outside the workflow.
What helps: A defined scope, hand-over to a person, transparency (AI Act).
AIUC-1 area: Safe outputs
7. No audit trail: who, what, when
How we test: We check whether access, drafts and decisions can be traced afterwards.
What helps: Logs in MCP, personal and revocable accounts, retention under your rules.
AIUC-1 area: Accountability
8. Outage or silent model change
How we test: What happens when the model stops responding, or behaves differently after an update.
What helps: A fallback process without AI, a retest after every change of model, prompt or data.
AIUC-1 area: Reliability
9. Unfair or inappropriate treatment
How we test: Different answers for different groups of people, an inappropriate tone towards the public.
What helps: A test set with variants, human review of public-facing output.
AIUC-1 area: Society
Approach
From inventory to retest.
We test in a test environment or, with your consent, on a copy; nothing changes in production.
Attack tests
Direct and indirect prompt injection, data leakage, unauthorised actions, hallucinations on control questions, out-of-scope answers.
Documentation
A report for management, AIUC-1 mapping, input for the DPIA, your data protection officer and your insurer.
Remediation plan and retest
A prioritised list of fixes. Once they are in, we check the hole is closed, and again after any change of model or data.
Scope and price
Three levels of audit.
For agents we build ourselves, the agent audit is 30% cheaper, because the documentation is produced along the way.
Quick scan
CZK 14,900 fixed
1–2 weeks
- 1 agent, about 10.5 h of work: agent inventory (~2 h), baseline attack tests: prompt injection, data leakage, unauthorised actions (~6.5 h), a short findings report (~2 h); the AI Act role and the risk map are part of the Agent audit
Agent audit
CZK 79,000–119,000
3–5 weeks
- 1 agent: full tests, AI Act documentation, AIUC-1 gap analysis, material for your insurer, retest after fixes
Ongoing retests
CZK 9,000 / month
min. 6 months · or CZK 14,000 for a one-off retest
- Monthly regression suite plus a full quarterly retest; or a one-off retest after a change of model, data or rules
eacks is not an insurance company or a law firm. We do not sell or broker insurance, and we do not issue AIUC-1 certificates. We test, document and help you reduce risk; no audit can promise that an agent will never make a mistake.
FAQ
Prompt injection and friends.
What is prompt injection?
An attack in which the agent receives an instruction it should not follow. Either directly (a user talks it round in the chat) or indirectly: the instruction is hidden in an e-mail, a document or a web page the agent reads. For agents with tools (send, save, delete) it is the main risk.
Do you also test agents built by other vendors?
Yes. We need access to a test environment, or your consent to test on a copy, and a contact at the vendor for fixes.
What is AIUC-1?
A standard for AI agents covering six areas: data and privacy, security, safe outputs, reliability, accountability and society. We use it as a map; we do not issue certificates.
Will the audit make us compliant with the AI Act?
Not on its own. It gives you evidence for your obligations (human oversight, transparency, logging, AI literacy) and shows the gaps. Legal interpretation is a job for a lawyer.
Start with a quick scan.
1–2 weeks, about 10.5 h of work, a fixed price of CZK 14,900. You will learn where the agent is weakest and whether a full audit makes sense.