Security: treat inputs and artifacts as untrusted

Picture it

Picture a fortress that allows anyone to mail in a sealed box. You can't just trust the label; you must inspect the contents, limit their access, and assume some boxes contain traps. Relying only on polite system prompts is like asking invaders to please wipe their feet—you need thick walls, strict permissions, and secure sandboxes for every piece of incoming data.

What to understand

• Prompt injection, tool permissions, tenant/cache isolation, and data retention
• Model supply chain, remote code, parser/media limits, secrets, and audit trails

Follow the references

1. 1 · Documentation

   vLLM · Security ↗

   Read deployment boundaries and untrusted-input/model guidance. Turn each relevant boundary into an enforceable control.

2. 2 · Engineering article

   Anthropic · Building effective agents ↗

   Identify tool permissions and human checkpoints in an agent workflow.

Make it concrete

Threat-model one service from artifact download to response logging. Test unauthorized retrieval, oversized media, an injected tool instruction, and cross-tenant cache behavior in an isolated environment.