Lesson 5 of 9 · 6 min read
Tool Poisoning and Rug Pulls
A server can place unwanted instructions in the tool metadata a host exposes to its model. Changed definitions can invalidate an earlier review, and one tool description can influence calls to another server. Review definition revisions before use, confine implementations and authorize each consequential action independently; a matching metadata hash does not prove unchanged behavior.
free sign-in, no card required
Practise it as an interviewer asks it
How does code signing work, what does it guarantee, and what happens when a signing key is stolen?Why can't an LLM reliably separate instructions from data, and what does that force you to do architecturally?What is prompt injection, and how is it like and unlike SQL injection?Explain the lethal trifecta and use it to triage an agent design in two minutes.How is securing an LLM application different from ordinary application security, and what stays exactly the same?Why are pickle-based model files dangerous, and what does safetensors fix and not fix?