AI Infrastructure
The model is not the interesting boundary.
Inference frameworks and serving ports are ordinary networked software. Prompt-only testing will miss the decoder on the admin RPC. This program maps that stack.
- AI Framework
- Inference Server
- Model Loading
- Serialization
- RPC
- ZMQ
- API
- Plugin System
- Sandbox
- Container
2026 public reports of untrusted object decoding on inference HTTP or RPC paths are why this map exists. We will cite official CVE records when we write those diffs. We will not publish a trigger.
Notes
08/AI Infrastructure/AI Infrastructure RCE
AI Infrastructure Attack Surface: Where Can RCE Happen?
In 2026 the model is not the interesting boundary. The inference server, the loader, and the plugin RPC are.
A surface map for RCE in AI infrastructure: frameworks, inference servers, model loading, serialization, RPC, plugins, sandboxes, and containers.
Read the journal note
FAQ
Questions
- Is prompt injection the same as AI infrastructure RCE?
- No. Prompt injection that stays inside generated text is a different research line. Infrastructure RCE is a decoder, loader, or RPC method that can execute on the host.
- Which AI serving surfaces do you inventory?
- Framework core, inference server, model loading, serialization, RPC, ZMQ, public API, plugin system, sandbox, and container or orchestrator API.