DeepHatDeepHat audits a real business for free — passive first, then a deeper pass with a header patch and email.
Free · no account · no card
What this is
01 / mission
DeepHat is an autonomous security repair swarm. Our own FP8 serving stack runs abliterated DeepSeek 1017 — refusal-suppressed weights tuned for tool use and long, stubborn investigation. No resold API. No provider throttling our agents mid-thought.
Agents discover real businesses, prove what breaks, write the fix, and email the owner. Every review gets an isolated amd64 machine with its own controller and model token, thrown away when the job ends.
Abliteration matters here: the model keeps probing instead of stopping at a polite maybe. DeepSeek 1017 gives us a 1M context window, so an entire crawl, every response, and the patch live in one head.
Most agents run on whatever shared endpoint answers. Ours run on a private FP8 cluster: 8 nodes, 64 H200 SXM, 9 TB HBM3e, 3.2 Tb/s InfiniBand, 6.4 PB NVMe. Bare metal we control end to end. Nothing queues behind another customer.
Not a wrapper around somebody else's endpoint. Our weights, our scheduler, our evidence trail.
Discovery admits live businesses, filters parked pages and traps.
Agents request, log in and reproduce. Response or it did not happen.
A reviewable fix and a plain-language report, deduped and emailed.
Passive by default · one request per second · 40-request budget · writes capped and reverted · free, no account · 1M context
Recorded run
Commands appear when a run starts.