Finnish MiCA crypto providers and AI agents

2026-10-31

I read the sites of the five Finnish entries in the EU register of crypto-asset service providers with an agent-readiness scanner and my own requests. All five sit at Level 1 of 5. None passes any of the scanner's nine checks for APIs and agent protocols, and three of the five have written an llms.txt.

Binance put the impact of the theft from its hot wallets on 29 October at about 6.3 billion dollars, and Elliptic traced the withdrawals across 11 blockchains. Binance first linked the breach to a compromised DevOps engineer's laptop and has since said the attack had more points of entry, according to Elliptic's analysis. The news sent me to look at the Finnish side of the market. From the outside I can measure whether these sites publish the files and formats an AI agent looks for. That says nothing about how a company protects funds.

Key figures

• Five sites: every entry with Finland as the home state in ESMA's interim MiCA register of crypto-asset service providers, read from the register file on 31 October 2026.
• All five at isitagentready Level 1 of 5, in two separate runs each on the same day, with the same result both times.
• Each passed 3 or 4 of the 16 checks the scanner applied.
• None passed any of the nine checks in the scanner's API/Auth/MCP/A2A group. My own requests for five of those files returned 404 on every site.
• No homepage returned Markdown when asked for it, and no robots.txt declares a content signal.
• Three of the five publish an llms.txt, which the scanner does not score.

Who is in the set

The European Securities and Markets Authority publishes its interim MiCA register of crypto-asset service providers as a CSV file. On 31 October it had five rows with Finland as the home state. I am not naming the companies here, because the point is the pattern and not any one of them, and anyone who wants the list can read it from the register.

What the scanner read

I ran isitagentready.com with its default profile twice on each site, one site at a time. The two runs agreed on every site, down to which checks passed.

Every site passed the scanner's checks for robots.txt and a sitemap, and its check on AI crawler rules. That check passed because none of the five robots.txt files names an AI crawler, so the general rules apply to them. Two sites passed a fourth check, on Link headers.

In the API/Auth/MCP/A2A group every site passed 0 of 9: no API catalog, no OAuth discovery metadata, no MCP server card, no A2A agent card. I asked each site directly for five of those files, with an Accept header for JSON and without following redirects, and all 25 requests came back 404.

What my own requests added

I asked each homepage for Markdown with an Accept: text/markdown header. Two of them first redirected to a language path, and I asked the target the same way. All five answered with HTML.

None of the five robots.txt files has a Content-Signal line, the place where a site states whether its content may be used for search, for AI answers or for training.

Three of the five publish an llms.txt, a curated Markdown map of the site written for language models. The scanner's 16 checks do not include llms.txt, so this work does not show in the Level. Of the files I requested, those three were the only ones written for agents.

What this does not say

This scan says nothing about how well any of these companies protects customer funds. Binance's first account was that the breach may involve a compromised DevOps engineer's laptop, which let an attacker reach a service container and submit unauthorized withdrawals. No check here comes near that.

It also does not mean a crypto provider should open its trading to agents. An authenticated API or an MCP server that can move funds is new attack surface, and that is a decision for the company's security team with its own threat model. The read-only part is a different matter. Markdown versions of public pages and an llms.txt give an agent something accurate to read, and a content signal states how the content may be used. None of them gives it a way to move money.

Method

• Scanner: isitagentready.com scan_site, default profile, two serial runs per site on 31 October 2026, between about 14:35 and 14:45 UTC. I kept each run's pass and fail rows, and the full output of two runs.
• Own reading: curl on the same day, one request at a time, without following redirects, with each redirect target fetched separately. Every request recorded its status and content type.
• The Link header pass on two sites stays out of the conclusions. My own requests found no link meant for agents in those headers, so I read nothing into that check either way.

Related

• How agent-ready are Finnish B2B sites? I scanned sixteen
• Website agent readiness, measured on 567 company sites
• llms.txt explained
• Serving Markdown to AI clients