Independent AI risk ratings and training

AI risk ratings and training for New Zealand small businesses.

Know where your business data goes when you use AI. We rate the tools you actually use, plan by plan, with a primary source for every claim. Then our courses train you and your team to use them safely, from a free certificate up to AI Risk Assessor.

AI tools rated

rated by plan, never averaged

primary sources, dated on every profile

Krebs on Security meets Money Saving Expert for AI risk. We rate vendors honestly; commercial relationships are disclosed openly. Every vendor profile carries a verification date and cites primary sources.

Start here

If you have not yet read our definition of the term, What is AI leakage? covers what the term means, including the three ways information leaks through AI tools (user error, vendor breach, attack surface), with worked examples from the canonical incidents of 2023-2026: Samsung, EchoLeak, CamoLeak, ForcedLeak, the Replit SaaStr database deletion, and the Otter.ai class action.

The Vendor Database

29 vendor profiles covering the AI tools that small businesses actually use. Each profile follows the same 12-section schema: risk rating, training defaults, retention, jurisdiction, breach history, and three honest takeaways for SMB owners.

The plan you are on changes the risk

Our risk rating for each of the 29 AI tools we track, from 1 (lowest risk) to 5 (highest risk). For 18 of them the rating depends on the plan: the open dot is the best plan we rate, the solid dot the worst.

Best plan we rateWorst plan we rate (or the only rating)
See the numbers
ToolBest planWorst plan
Grok3 of 55 of 5
Otter.ai3 of 55 of 5
Replit3 of 55 of 5
Midjourney3 of 54 of 5
ElevenLabs2 of 54 of 5
GitHub Copilot2 of 54 of 5
Granola2 of 54 of 5
Jasper3 of 53 of 5
Shopify Magic / Sidekick3 of 53 of 5
Slack AI3 of 53 of 5
ChatGPT2 of 53 of 5
Claude2 of 53 of 5
Fathom2 of 53 of 5
Grammarly2 of 53 of 5
Microsoft Copilot2 of 53 of 5
Perplexity2 of 53 of 5
Canva Magic Studio1 of 53 of 5
Cursor1 of 53 of 5
Gemini1 of 53 of 5
Asana AI / Asana Intelligence2 of 52 of 5
HubSpot AI2 of 52 of 5
Intercom Fin2 of 52 of 5
Linear AI2 of 52 of 5
Salesforce Einstein / Agentforce2 of 52 of 5
Xero2 of 52 of 5
Zendesk AI2 of 52 of 5
ZoomMate2 of 52 of 5
Adobe Firefly1 of 52 of 5
Notion AI1 of 52 of 5

Source: AI Leakage vendor profiles, each rating dated and sourced on its own page. How we rate: methodology.

By category

Direct AI Tools

12 profiles

ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity, Grok, GitHub Copilot, Cursor, Replit, Jasper, Midjourney, ElevenLabs

Embedded Productivity AI

8 profiles

Slack AI, Notion AI, ZoomMate, Asana AI, Linear AI, Granola, Otter.ai, Fathom

Vertical SaaS AI

9 profiles

Intercom Fin, Zendesk AI, Salesforce Einstein/Agentforce, HubSpot AI (Breeze), Shopify Magic, Canva Magic Studio, Adobe Firefly, Grammarly, Xero/JAX

Changes that mattered, month by month

We watch 147 pages for the tools we track and for New Zealand privacy law: terms, privacy, pricing and policy pages, most of them checked daily. Since monitoring began we have logged 971 page changes. Each bar counts the changes we reviewed and found to matter to a business: 56 so far.

See the numbers
MonthPage changes loggedChanges that mattered
August 20261576
September 202628910
October 202635430
November 2026 (so far)17110

Source: AI Leakage change monitor, as at 19 November 2026. Monitoring began on 3 August 2026, so August is a part month.

What we cover that other sites don’t

  • SMB-shaped recommendations. The privacy press tends to focus on either consumer privacy or enterprise compliance. AI Leakage is written for the 1-to-10 employee business segment. Recommendations are sized for that scale.
  • Tier-specific risk ratings. Many AI tools have materially different privacy postures between consumer and enterprise tiers. We rate each tier honestly rather than averaging.
  • Threat patterns, not incident lists. Most privacy advice tells you what to do. It rarely tells you when the advice stopped working. We track how a class of risk moves, and which of the controls you rely on it has already outrun. Two sets, deliberately kept apart: how AI vendors’ practices evolve, and how AI is used against your business: faked invoices, cloned voices, forged identity documents. Each pattern says which industries it lands on hardest, and shows the working.
  • Active litigation tracking. Vendors in the database are defending class actions that go directly to how their products handle your data (Otter.ai and Granola). We record the case number, the court and what is actually pleaded, and where a date we previously published can no longer be confirmed from the docket, we say so rather than leaving it up.
  • NZ-relevant context. The site is operated from New Zealand. NZ Privacy Act 2020 considerations and NZ-relevant vendors (Xero JAX) get explicit treatment.

How this site uses AI

AI Leakage uses Claude (Anthropic) in its content workflow. We disclose this on every relevant page. Claude is also rated here, on the same method and evidence standard as every other vendor. Full disclosure here.

Editorial principles

  • Independence of verdicts. No vendor pays for a favourable rating, and we take no affiliate commission on any product, rated or not. We do not accept sponsored evaluations. Where another commercial relationship exists, it is disclosed openly on the affected page.
  • Every factual claim sourced. Every cited URL verified on the date shown at the top of each profile.
  • If our recommendation is uncomfortable, we make it anyway.

Full methodology including the risk rating scale, the 12-section profile schema, and the commercial relationships we will and won’t accept is on the Methodology page. Current affiliate relationships are listed on the Affiliate Disclosure page.

For practitioners

Bookkeepers, IT providers, consultants and others who advise businesses on AI can take the Advanced certificate and train their own staff and clients at a lower practitioner rate. We pay nobody for introductions. See For practitioners.

Get in touch

Corrections, tips, vendor disclosures, research collaboration: send us a message with the contact form. We respond to factual corrections within five business days.