Method
How the AI control finder works
Short answer
By The Device Desk · Endpoint AI Finder · Published 2026-10-11
Desk research from public vendor pages read on 27 September and 11 October 2026. Profiles follow our published method; we did not run the products.
Vendor pages read September and October 2026 · updated 2026-10-13
What do we record for each product?
For each of the four product control types (on-device control, browser extension, network or SSE proxy, AI or MCP gateway) we record one of three statuses with a note and a source: Described (the product pages we read describe it), Partly described (it covers part of the type, sits in another component, or the method is not detailed), and Not in our sources (we did not find it). Policy only is the fifth type and needs no product, so it is not recorded per product. We also record ten checks: naming the coding agents covered, deciding on a tool call before it runs, controlling what goes into browser AI, rolling out through EDR or device management, warning or asking instead of only blocking, blocking, a watch-only or simulation mode, recording who or what acted, covering SaaS, cloud or homegrown AI, and publishing pricing or tiers. Each check is Matches, Partly or Not in our sources.
- DEV Filled marker: described on the product pages we read
- DEV Hatched marker: partly described (another component, a narrower scope, or method not detailed)
- DEV Dashed marker: not in our sources (it may exist; we did not find it)
How does the finder weigh the control types?
| Question | Answer | Points |
|---|---|---|
| 01Which AI tools do employees use at work? | AI chat and assistants in the browser | BRW +3, NET +1 |
| 01Which AI tools do employees use at work? | Desktop AI apps and agents | DEV +3 |
| 01Which AI tools do employees use at work? | Coding agents and MCP servers | DEV +4, GWY +2 |
| 01Which AI tools do employees use at work? | Not sure yet | DEV +1, BRW +1, POL +2 |
| 02Do you manage the devices people use for AI? | Yes, company-managed devices | DEV +2 |
| 02Do you manage the devices people use for AI? | A mix of managed and unmanaged | BRW +2, GWY +1 |
| 02Do you manage the devices people use for AI? | Mostly unmanaged devices | BRW +2, GWY +2, POL +2 |
| 03Where does the work happen? | Mostly on laptops and desktops | DEV +1 |
| 03Where does the work happen? | Partly in cloud workspaces | GWY +1 |
| 03Where does the work happen? | Mostly in cloud workspaces or virtual desktops | NET +1, GWY +3 |
| 04How sensitive is the data AI tools can reach? | Mostly public or general internal information | POL +3 |
| 04How sensitive is the data AI tools can reach? | Customer data or source code | DEV +1, BRW +1 |
| 04How sensitive is the data AI tools can reach? | Regulated data, secrets or production access | DEV +2, NET +1, GWY +1 |
| 05Which of these do you already run? | Any of Palo Alto Networks, SentinelOne, CrowdStrike, another EDR or device management (once, however many are picked) | DEV +2 |
| 05Which of these do you already run? | An SSE or secure web gateway | NET +4 |
| 05Which of these do you already run? | A DLP product | BRW +1 |
| 05Which of these do you already run? | None of these | POL +1 |
| 06When an AI action looks risky, what do you want to happen? | Block it | DEV +2, NET +1 |
| 06When an AI action looks risky, what do you want to happen? | Warn or ask the person | DEV +1, BRW +2 |
| 06When an AI action looks risky, what do you want to happen? | Only watch for now | DEV +1, POL +2 |
| 07How widely is AI used? | A handful of people | POL +3 |
| 07How widely is AI used? | One or two teams | Nothing |
| 07How widely is AI used? | Across most of the company | DEV +1, BRW +1 |
Every type starts at 0. Types are ordered by points, highest first; ties keep the fixed order DEV, BRW, NET, GWY, POL. The first is Start here. Of the next two in that order, each is Also relevant if its points are at least 3, or half the first type's points if that is higher. The rest with points above 0 are Lower on your list; a type with 0 points is Not suggested by your answers.
If Start here is policy only, the result opens with "You may not need a product yet", a short checklist of policy steps, and then the product list under "If you still want to look at products", checked against the other picked types. If policy only is Also relevant, a note says that a product does not replace an approved-tools list and clear rules.
How does the finder check products?
Each product gets points for every picked product control type (Start here and Also relevant, without policy only): Described 2, Partly described 1, Not in our sources 0. Then for each check your answers trigger, and for what you already run:
| Check | When it applies | Matches | Partly | Not in our sources |
|---|---|---|---|---|
| Names the coding agents it covers | AI tools include coding agents and MCP servers | 2 | 1 | 0 |
| Decides on a tool call before it runs | AI tools include coding agents and MCP servers, or desktop AI apps and agents | 2 | 1 | 0 |
| Controls what goes into browser AI | AI tools include browser AI | 2 | 1 | 0 |
| Rolls out through EDR or device management you run | Devices are not mostly unmanaged and you run Palo Alto Networks, SentinelOne, CrowdStrike, another EDR or device management | 2 | 1 | 0 |
| Records who or what took each action | Data is regulated data, secrets or production access | 2 | 1 | 0 |
| Also covers SaaS, cloud or homegrown AI | Work happens partly or mostly in cloud workspaces | 2 | 1 | 0 |
| Can block an agent, tool or action | You want to block risky actions | 2 | 1 | 0 |
| Can warn or ask instead of only blocking | You want to warn or ask the person | 2 | 1 | 0 |
| Has a watch-only or simulation mode | You want to only watch for now | 2 | 1 | 0 |
| Has a public price or tiers (vendor site or marketplace) | A handful of people use AI | 2 | 1 | 0 |
| Built into a platform you already run | For each platform you picked that the product is part of | +4 | ||
| Works with something you already run | For each tool you picked that the product's pages say it works with | +2 | ||
Products with the same total are shown as equals and listed alphabetically. We never show the total as a score.
What does Not in our sources mean, and is it fair?
It means we did not find the item on the pages we read. It counts zero, not against the product. It does affect order: a product whose pages describe a simulation mode can match a watch-only check, and one whose pages do not cannot. Younger and quieter vendors publish less, and that can cost them here. Where this cuts against a product that does more than its pages say, the fix is a source, not a guess: send it through the contact page and we will add it.
Which products are covered, and why these 19?
19 products that control or watch AI agents, coding assistants, MCP servers or browser AI for employees, or the AI traffic of apps a company builds, with public pages detailed enough to record: AI Access Security (Palo Alto Networks), Akamai Workforce Protector, Bay, Bloom Security, Cloudflare AI Gateway, CrowdStrike Falcon Guardian, Ent, Glow Security, Harmonic Security, Koi (Palo Alto Networks), Lasso Security, Neo, Noma Security, Onyx Security, Pluto Security, Prompt Security (SentinelOne), Sun Security, Zenity and Zscaler Generative AI Security. Four of them sit on the network or at a gateway rather than on the device: AI Access Security (Palo Alto Networks) and Zscaler Generative AI Security in the network path, Cloudflare AI Gateway in front of AI applications, and Akamai Workforce Protector in the browser. We also read pages for Island, an enterprise browser that describes itself as The Enterprise Agentic Control Plane, and Aim Security, which Cato Networks acquired on 3 September 2025. We could not read enough of either, or of Netskope's AI pages, to record them yet.
Source: Island homepage · Read 27 Sep 2026
Source: Aim Security homepage · Read 27 Sep 2026
Source: Cato Networks press release: Cato acquires Aim Security · Read 11 Oct 2026
Where do the facts come from?
- Vendor product pages, documentation, pricing pages, marketplace listings, blog posts and press releases, read on 27 September 2026 and, for later additions and corrections, on 11 October 2026. Each fact carries its source link and its own read date. A few facts come from trade press and are marked as secondary sources.
- Vendor figures such as deployment times, false-positive rates or the number of rules are quoted as the vendor's statement and are never our finding.
- Where a vendor does not publish a price, we write Not published. Only marketplace listings show prices: AWS Marketplace lists the Noma Security Platform at $200,000 for 12 months and Lasso Security's chatbot product at $20,000 for 12 months. No vendor site in this set publishes a list price.
- Model Context Protocol terms link to the MCP specification and its security best practices; OWASP terms link to the OWASP GenAI Security Project.
What are the limitations?
Public sources only. No hands-on testing, no demos, no vendor interviews. A vendor's pages describe what it sells, not how well it works on your devices. Products change fast, and several sit inside larger companies (Koi and AI Access Security at Palo Alto Networks, Prompt Security at SentinelOne, Workforce Protector at Akamai), whose pages may lag the product. Use the finder to decide what to ask and what to put into a pilot, not as a test result.
Does any product always come first?
No. The same rules apply to every product, they are published on this page, products with equal fit are listed alphabetically, and Bay's product notes list its limits, including that it covers the endpoint only, that it has one of the smallest public footprints here with no published pricing, and that its headline figures are its own statements. With other answers, other products come first: a company starting with browser AI and coaching is pointed to Harmonic Security, a Palo Alto Networks shop to Koi, a SentinelOne shop to Prompt Security, a company behind an SSE with cloud workspaces to Lasso Security, and the default example to Lasso Security and Onyx Security as equals. Two of our eight example profiles point to policy before any product.
How do we handle corrections?
Send a correction through the contact page with the topic Correction and a public source. We check it against the vendor's page and update the data; finder results recompute automatically.