Changelog
Documentation
Roadmap

March 24th, 2026

Improved

Capability

Julius: 30 New LLM Service Probes — From Ollama to AWS Bedrock

Your Attack Surface Now Includes LLM Infrastructure. Julius Finds It.

Julius nearly doubles its detection coverage with 30 new probes, bringing the total to 63 LLM services across self-hosted engines, cloud APIs, gateways, and RAG platforms. Every new probe was live-tested against real instances via Shodan and customer infrastructure — and the testing uncovered 14 detection bugs in existing probes that are now fixed.

New Service Detection (30 probes)

Cloud-Managed APIs (10)

Julius can now fingerprint major cloud AI providers without valid credentials.

  • AWS Bedrock — Detects both control plane and runtime endpoints via 403 response
  • Azure OpenAI — Microsoft Foundry deployment detection
  • Google Vertex AI — aiplatform.googleapis.com endpoint fingerprinting
  • Groq, Fireworks AI, Together AI, Replicate, Modal — Provider-specific error signatures
  • Cloudflare AI Gateway — Detected via unique cf-aig-event-id response header
  • Databricks Model Serving — Real-time ML inference endpoint detection

Self-Hosted Engines (10)

  • SGLang — High-performance LLM serving via /server_info and /model_info
  • NVIDIA TensorRT-LLM, Triton Inference Server — Enterprise GPU inference
  • BentoML, Baseten Truss, Ray Serve — ML serving frameworks
  • DeepSpeed-MII, MLC LLM, Petals, PowerInfer — Specialized inference engines

Gateways (5)

  • Bifrost, Helicone, OmniRoute, Portkey AI Gateway, TensorZero — LLM proxy and routing layers

RAG and Orchestration (5)

  • Langflow, h2oGPT, PrivateGPT, Quivr, RAGFlow — Document QA and workflow platforms

Bug Fixes From Live Testing (14 probes fixed)

Cross-Version Compatibility

  • SGLang — Fixed /server_info match for older versions
  • Dify — Newer JS-rendered versions now detected via data attributes
  • Flowise — Updated title match for current versions

False Positive Elimination

  • Ollama — SGLang and Ollama-compat servers no longer trigger false matches
  • Bifrost — Removed overly generic /api/version block
  • DeepSpeed-MII — Removed /health block that matched every FastAPI app
  • Groq — Removed generic block that matched KoboldCpp

Systemic header.contains Bug (5 probes)

AWS Bedrock, Cloudflare AI Gateway, Fireworks AI, Modal, and OmniRoute all had header matching rules that never worked on HTTP/2. All five are now fixed.

Get Started

go install github.com/praetorian-inc/julius/cmd/julius@latestjulius probe https://your-target.example.com

Julius is open source at github.com/praetorian-inc/julius.

Write a comment...