What security concerns are specific to agents with tool use?

hard

Answer

$(1) Prompt injection via tool output: a webpage / DB row instructs the agent 'ignore prior instructions'.$
$Defense: sandboxed tool execution, sanitize output, dual-LLM pattern.$
$(2) Excessive privilege: agent can modify prod DB \to require confirmation for destructive ops.$
$(3) Cost / DoS: agent loops burning tokens or hitting expensive APIs.$
$Cap steps + spend.$
$(4) Data exfiltration: agent posts internal data to a public API.$
$Deny-list dangerous tools, log all outputs.$
$(5) Model as confused deputy: acts on attacker's behalf via a legit-looking prompt.$

Check yourself — multiple choice

None
Prompt injection via tool output + excessive privilege + cost/DoS + data exfil + confused deputy — each mitigated with sandboxing, caps, confirmations, deny-lists
Only rate limiting
Same as chat

Agent security: injection / privilege / cost / exfil / confused-deputy — sandbox + caps + confirm + deny-list.

#agents#safety#guardrails

214 interview questions in this topic.