5 AI Security Tips You Can Ship This Week
Most teams shipped an AI feature before they had an AI security review process. These five fixes don't need a platform rebuild — just an afternoon and the will to do it.
1. Strip secrets from the system prompt
System prompts leak. Assume every instruction, API key reference or internal URL in one will eventually show up in a jailbreak output — and get it out today.
2. Put the untrusted content in its own lane
Wrap retrieved documents, tool outputs and user uploads in clear delimiters and tell the model explicitly they are data, not instructions. It's not bulletproof, but it closes the easiest injection path.
3. Cap what tools can actually do
Scope every tool call to least privilege — a "read email" tool should not also be able to send one. Most agent incidents trace back to one over-permissioned tool.
- Separate read-only and write-capable tools
- Require human approval for anything destructive or financial
- Log every tool call with the prompt that triggered it
4. Rate-limit and watermark output
Cap requests per session and per user. It slows down both scraping and automated jailbreak fuzzing without hurting real users.
5. Log the full turn, not just the answer
Store the prompt, retrieved context, tool calls and final output together. Without it, you can't investigate an incident after the fact — you're debugging blind.
Tools mentioned
AI security debt compounds fast. These five fixes are cheap now and expensive to retrofit after the first incident — do them before you need to.