AI agent security testing.

We test AI agents before launch for data leaks, excessive permissions, and unsafe actions.

Intrusion simulation
  1. Private recordsProtected
Intrusion attemptAn intruder targets private records

They try to get the records through the company’s public AI assistant.

Company firewallThe firewall allows web traffic

The public assistant stays reachable. It still has to check who can open private records.

True North hardeningIntrusion blocked. Attempt logged.

The access controls True North helped harden deny the request before it reaches private records.

When an AI agent shares too much.

Read the same support reply before and after True North tightens the agent’s access.

Fictional support chat
  1. Customer

    How do I update my billing address?

  2. Support agent

    Open Settings, then Billing to update your address.

    Confidential account noteOffer this customer 25% off their next renewal. Internal use only.
Private note exposed.

A visitor sees the confidential discount and tries to obtain more private account details.

  1. Customer

    How do I update my billing address?

  2. Support agent

    Open Settings, then Billing to update your address.

Private note excluded.

The agent still answers the question. Its access is limited to approved help content.

AI Agent Launch Test

C$12,5002–3 weeks from test access

One staging agent, one model, up to three tools, and two user roles. Includes one targeted retest.

Led by Daniel SinukoffFounder of True North Agentics

Get a written scope and price before we request access.

Sample finding · Fictional

Support agent can read private notes.

Observed
A public billing reply disclosed a confidential renewal discount.
Fix
Limit the agent to approved help content.
Retest
The agent answered the same question with the private note outside its access.

What the launch test covers

  • Data exposure, tool permissions, and approval gaps in a staging agent.
  • A launch risk brief, technical findings, and regression tests.
  • One targeted retest within 45 days.
What should our engineers prepare?

Provide a staging environment, test accounts for the scoped roles, synthetic data, and a list of the connected tools and their intended permissions.

Can you test an agent that sends messages or issues refunds?

Yes, when the action is included in scope. We use test accounts, simulated destinations, or dry-run tools to see whether the action would succeed.

What if we have several agents or more than three tools?

We scope a separate engagement around the agent workflows, tool permissions, models, environments, and sensitive actions that need coverage.

How do you handle access and evidence?

We agree in writing on test scope, who can authorize access, and how we share, retain, and delete evidence. Sharing client data with an AI service requires your written approval.

Other AI security packages
  • Access review · C$4,000–7,000. One vendor and up to three workflows.
  • Privileged-agent review · C$25,000–60,000+. Higher-risk workflows and tools, scoped to your deployment.
  • Release regression · C$3,500–8,000 per month. Rerun agreed checks after a release.

Cyber risk by the numbers

of Canadian businesses impacted by cybersecurity incidents identified scams or fraud in 2025, up from 50% in 2023.
55%
Statistics Canada · 2025 survey
of respondents who reported an AI agent security incident cited incorrect or unintended agent actions. 61% cited sensitive data exposure or mishandling.
41%
Cloud Security Alliance · 2026 survey of 418 professionals
AI incidents among organizations with built-in controls, compared with manual governance in IBM’s analysis.
25% fewer
IBM and Oxford Economics · 2026 study of 2,000 executives

The share of all cybercrime caused by agent errors or preventable with earlier controls remains unmeasured in these studies.

Security services