We trained a small model to guess subdomains
Why we built our own 18-million-parameter model for subdomain enumeration, whether ChatGPT would do it better, how many hidden names it finds and what made no difference.
FIELD NOTES
Investigation methods, infrastructure research and engineering notes from the people building Profundis.
Why we built our own 18-million-parameter model for subdomain enumeration, whether ChatGPT would do it better, how many hidden names it finds and what made no difference.
4,500 Danfoss interfaces on Shodan: how many fridges? We examined the count, then the names associated with controllers in our data.
4 500 interfaces Danfoss sur Shodan : combien de frigos ? On a examiné le décompte, puis les noms associés aux boîtiers dans nos données.
Mail and DNS providers of the top 1M domains, from the MX and NS records of 130,803 domains. Google Workspace 22.7%, Microsoft 365 18.0%, Cloudflare 34.6%.
I took one junk email from my quarantine folder, pivoted on its WHOIS registrant with Claude Fable 5 and the Profundis MCP, and ended up on the web server of a registered French webmarketing company.
What if you could search for similar favicons, not just exact hash matches? Discover how we built a favicon similarity search that processes 50+ million images in seconds.
Because piping commands together is not the answer. Learn how we built an infrastructure that processes billions of public assets.