I scanned 39 AI companies' DNS records — here's who's verified with Anthropic, who's deploying MCP keys, and whose email you can spoof

DNS records are public. When a company verifies domain ownership with a service provider, that verification lives in their TXT records — visible to anyone who runs dig TXT example.com. These records form a map of vendor relationships that companies may not realize they're broadcasting.

I queried the DNS, SSL, WHOIS, and HTTP headers for 39 AI-adjacent domains — labs, safety orgs, tooling companies, prediction markets. Here's what's actually in there, verified with independent dig queries.

Anthropic domain verification: who's completed it

17 out of 39 domains contain an anthropic-domain-verification TXT record:

What this means: these companies completed Anthropic's domain verification process — likely part of Claude API enterprise onboarding. It doesn't prove an enterprise contract, but someone at these organizations generated a verification token and added it to their DNS.

Notable absences:

Cursor verification shows a similar pattern. 14 domains have cursor-domain-verification records, including Anthropic itself. Companies like Hugging Face, Modal, Fireworks, Pinecone, Vercel, Sourcegraph, and Stripe appear in both Anthropic's and Cursor's verification lists.

MCPv1 DNS records: cryptographic MCP server identity

This one surprised me. Multiple companies have v=MCPv1 TXT records containing public keys:

The record format (v=MCPv1; k=<algorithm>; p=<public-key>) resembles DKIM's DNS key structure. This appears to be a DNS-based identity verification mechanism for MCP (Model Context Protocol) servers — allowing clients to verify that an MCP endpoint is actually operated by the domain owner.

I couldn't find a public specification for this record format. Microsoft's investment of 12 separate key pairs suggests multiple verified MCP services. The ecosystem is early — only 6 of 39 domains have these records.

Email security: who's vulnerable to spoofing

DMARC tells receiving mail servers what to do with messages that fail authentication. p=reject blocks spoofed messages. p=none lets them through.

That's 9 out of 39 domains (23%) with weak or absent email authentication. For AI safety organizations handling sensitive research — MIRI and Alignment Forum — this is more concerning, given the sophistication of potential adversaries.

The strong side: Anthropic (p=reject), OpenAI (p=reject), Stripe (p=reject), and Cloudflare (p=reject) all have strict policies.

SPF softfail: the policy that doesn't protect

HN user sybercecurity asked a good question: how many companies use ~all (softfail) in their SPF records? Softfail tells receiving servers "this message probably isn't from us, but deliver it anyway." It's the default for many email setups, and it barely enforces anything.

Two-thirds of AI companies use softfail. Including Anthropic, Google, Apple, NVIDIA, Hugging Face, Perplexity, Scale AI, Cohere, Stability AI, and Mistral. Only OpenAI, Microsoft, Amazon, Palantir, x.ai, Pinecone, Jasper, and Runway use hardfail (-all).

Three domains have no SPF at all: Meta (relies on other mechanisms), Tesla, and Alignment Forum.

The practical implication: if you receive an email "from" most AI companies, your mail server won't definitively reject it even if it's spoofed. The ~all softfail tells your server to accept it but maybe flag it. Combined with weak DMARC (as we saw above), this means email impersonation is technically feasible against the majority of companies in this dataset.

Check any domain's email security

Enter a domain below for a combined spoofability assessment. Unlike checking SPF or DMARC in isolation, this weighs all three layers together — SPF policy, DMARC enforcement, and DKIM presence — because a domain with SPF softfail but strict DMARC + DKIM alignment is still well-protected. Data pulled live from public DNS records via DomainIntel.

Smaller findings

Methodology

Data collected May 20, 2026. All sources are public — DNS records, SSL certificates (Certificate Transparency), WHOIS registration, HTTP headers. To verify any claim:

Data was collected using DomainIntel, which combines DNS, SSL, WHOIS, and tech detection into one call. During this analysis I found and fixed two bugs in its DMARC detection. All DMARC claims were independently verified with dig.