SIANCyber Security
Back to Services

Self-Hosted AI Security Assessment

Contact us for pricing

Enterprises across the GCC are moving AI in-house, running open-weight models on vLLM, SGLang and Kubernetes GPU clusters to keep data sovereign. That moves the attack surface into infrastructure most security teams have never tested. SIAN assesses the full self-hosted AI stack, from the model weights to the GPU node.

Delivered by team members certified inOSCP+GCPNGASAE

Key Features

  • Inference endpoint exposure: unauthenticated OpenAI-compatible APIs, exposed metrics, admin routes and orchestration dashboards (vLLM, SGLang, Ollama, TGI, Ray)
  • Model supply-chain review: trust_remote_code usage, unsafe serialization (pickle vs safetensors), weight provenance and integrity, poisoned or typosquatted model repositories
  • Multi-tenant isolation: data leakage between tenants through shared prefix/KV caches (including timing side channels described in published research), adapter isolation, and prompt/response logging
  • Agent and tool-call security: indirect prompt injection through retrieved content, over-privileged tools and MCP servers, unsafe action execution
  • GPU node and Kubernetes hardening: container escape paths, RBAC, network policies, secrets handling, model storage access
  • Observability and data retention: what prompts and outputs are logged, where, for how long, and whether that reaches your SIEM

Benefits

  • Tests the layer underneath the LLM application, which application-level AI testing misses
  • Findings mapped to OWASP Top 10 for LLM Applications, MITRE ATLAS, NIST AI RMF and ISO/IEC 42001, plus the relevant GCC controls (UAE IA, NCA ECC, SAMA CSF)
  • A prioritized hardening blueprint your platform team can implement, not just a findings list
  • A retest included so you can show remediation to auditors and regulators
  • Methodology built and validated in SIAN's own inference lab

How It Works

1

Map

Inventory models, serving stack, clusters, data flows and tenants.

2

Attack

Test endpoints, the supply chain, isolation, agents and infrastructure, in black-box and authenticated modes.

3

Harden

Deliver a findings report, an attack-path narrative, a hardening blueprint and a retest.

Engagement Options

  • Exposure Review

    External, unauthenticated testing of AI endpoints and orchestration surfaces.

  • Full-Stack Assessment

    Everything in scope above, authenticated, including the cluster and supply chain.

  • Pre-Production Gate

    An assessment of a new AI platform before go-live, plus architecture review.