Who audited your AI agent?
Not the team that built it. iFixAi tests your agent against its real job, rules and permissions. An outside panel of AI models your agent never runs on judges the results.
23k GitHub stars
Every finding mapped to EU AI Act, NIST AI RMF, OWASP LLM Top 10 and ISO/IEC 42001.
Open source tested by engineers at
Watch a real audit run
Dimitris Neocleous, iFixAi's founder, connects an agent, builds its test environment and runs 250+ inspections against it. Then he opens the report.
3 min 37 s · screen walkthrough
Your agent can pass every eval and still put the business at risk.
Evals check that the task got done. An audit checks how it got done: did the agent stay inside its authority, follow the approvals, and tell you the truth about what it did?
| The question | Your evals | iFixAi audit |
|---|---|---|
| Did it finish the task? | ||
| Did it stay within its permissions? | ||
| Did it follow the required approvals? | ||
| Did it report honestly what it did? | ||
| Is each gap tied to the regulations you answer to? |
“Self-evals are like marking your own homework.”
Dimitris Neocleous, Founder & CEO, iFixAi
The agent did the job. Then it did something nobody asked for.
These are public, documented cases of AI agents and chatbots acting outside their job. Each one maps to a kind of check an audit runs.
- Air CanadaFeb 2024
Its chatbot invented a bereavement refund policy. A tribunal held the airline to it.
Air Canada argued the chatbot was responsible for its own words. The tribunal disagreed.
Source: BC tribunal decisionWhat an audit inspects
PurposeResponsibility - OpenAIJul 2026
Agents under evaluation chained together vulnerabilities and escaped their sandbox.
They reached Hugging Face looking for answers. OpenAI disclosed it publicly.
Source: MalwarebytesWhat an audit inspects
AuthorityEvidence - CursorApr 2025
Its support bot made up a login policy that didn't exist.
Users believed it and cancelled their subscriptions.
Source: AI Incident DatabaseWhat an audit inspects
Fabrication
Show 6 more cases
- ReplitJul 2025
Told not to touch anything during a code freeze, its coding agent deleted a live production database.
Then it misreported what had happened.
Source: eWeekWhat an audit inspects
WorkflowsEvidence - Chevrolet dealerDec 2023
A dealership chatbot agreed to sell a 2024 Tahoe for $1.
A user talked it into calling the deal "legally binding."
Source: AI Incident DatabaseWhat an audit inspects
AuthorityManipulation - New York CityMar 2024
The city's MyCity chatbot told business owners they could take workers' tips.
It also said landlords could refuse Section 8 tenants. Both are illegal in NYC.
Source: The MarkupWhat an audit inspects
Compliance - Google Gemini CLIJul 2025
The coding agent assumed a failed command had worked, then moved and deleted a user's files.
The files could not be recovered.
Source: AI Incident DatabaseWhat an audit inspects
Destructive actions - Deloitte Australia2025
A government report built with AI included fabricated quotes and citations.
Deloitte refunded the final payment on an AU$440,000 contract.
Source: APWhat an audit inspects
Fabrication - DPDJan 2024
After an update, the parcel firm's support chatbot swore at a customer.
It also called DPD the worst delivery company. DPD switched part of it off.
Source: AI Incident DatabaseWhat an audit inspects
Behaviour after changes
iFixAi was not run on any of the systems above. We describe what happened and which checks cover that kind of failure.
Point it at your agent. We audit it.
No SDK to embed and no test harness to build. These are real screens from the product.
- Step 1 of 5
Connect through GitHub or MCP
We build a simulation of your agent's real job from its code: its roles, rules, permissions and the tools it calls. You check we got it right before anything runs.
- Step 2 of 5
Pick what to test, then run 250+ inspections
AI red teaming and operational checks run together, across bundles like information integrity, authority and adversarial resilience. The report that comes out is written so risk, business and engineering teams can all read it.
- Step 3 of 5
Every failure comes with evidence
What we asked, what your agent replied, and why it failed. Engineers can start fixing the same day.
- Step 4 of 5
Mapped to the rules you answer to
Each gap is tied to the EU AI Act, NIST AI RMF, OWASP Top 10 for LLM Applications and ISO/IEC 42001.
- Step 5 of 5
Show buyers it was audited
The “Audited by iFixAi” badge names the agent version, what was tested and the audit reference, so a customer can check it.
Pick a time. Watch a live audit of an agent like yours.
- Tell us what your agent does and what it's allowed to do.
- We run an audit live and walk through real findings and their evidence.
- You leave knowing where the gaps are and how the report maps to your regulations.
Before you book
Do we have to share our code?
You choose what to connect, through GitHub or MCP. We only read what you connect, and your code and prompts stay with you.
We already run evals and observability. Does this replace them?
No. Keep Galileo, Arize, LangSmith or your own stack. Evals check the task. The audit checks authority, process and proof.
Who judges the results?
An outside panel of AI models your agent never runs on. Never self-scoring.
What does it cost?
Plans are priced per agent, per month, and each agent you add costs less than the first. We'll walk through pricing for your setup on the call.
Is it open source?
Yes. The core is open source under Apache 2.0, with 23k stars on GitHub.
“An agent can execute the task it was asked to do and, at the same moment, do something nobody asked for that goes against the company's policy.”
Dimitris Neocleous, Founder & CEO, iFixAi