AI Security
AI security here means the places where a language model or an agent meets something that can act: a shell, a file system, a database, a browser, or a package registry. The notes keep what the model says separate from what the surrounding code allows.
The path starts with model output treated as an input like any other, moves to agents that read data which can contain instructions, then lists the permission settings of three coding agents with the versions checked, and ends with a decision guide that ran two of those agents on a harmless task. The first two experiments call no model; each note says what its method does and does not show.
Reading order
Treat model output as untrusted input
Start here: four ways that using a model's output directly turns text into an action, and the fix for each.
When the data an AI agent reads can give orders
What changes when an agent reads outside data and then acts, shown with a local simulation.
AI coding agents: permission and safety settings
Reference: how three coding agents ask for approval and limit where they can act, with the versions checked.
Claude Code or Codex: which permission setup fits the task?
Decide: the same four harmless steps run through Claude Code and Codex, what each configuration did, and what the test could not show.
What we have not covered yet
- How real models behave under attack; the only live agent runs were four harmless file and network steps.
- Evaluating model quality or safety.
- Retrieval-augmented generation and model supply chain security.