Safe by Default: Building Agents That Can't Wreck Your Data
Read/write enforcement, principal scoping, approval gates, sandboxes and budgets: five safety properties, every one enforced in code, not in a prompt.
Agent Harness03
Subject · 4 posts
prompt injection
Exfiltration Through the Model's Mouth: Output Is an Attack Surface
The model's output came from your trusted system, so it feels safe. It is not: untrusted input shaped it, and one rendered image can carry your data to the attacker.
LLM and Agent Security05
Prompt Injection: Direct, Indirect, and Why Filtering Fails
The signature attack on language models. How instructions hide in the text a model reads, why you cannot filter your way out, and what containment looks like.
LLM and Agent Security02
Why AI Security Is a New Problem: The Model Has No Data Channel
Every security bug in the last series was data mistaken for code. A language model makes that its whole way of working: to it, all text is instructions.
LLM and Agent Security01