← All categories
Friday, 20 November 2026Today · official
CybersecurityDifficulty ★★★★☆2026-11-20

AI agent tool calls slip past intended safety boundaries

Your team has deployed a customer-service AI agent for an online retailer. The agent is backed by an LLM and can invoke a suite of backend tools — issuing refunds, applying discounts, and looking up account records. Within two weeks of launch, your security team flags three incidents: a refund that exceeded the authorised cap, a discount applied to a product category that was supposed to be excluded, and a data lookup that crossed customer session boundaries. The LLM logs show the model issued each call with apparent confidence; no error was raised at the tool layer. Engineering has added ad-hoc guards in several tools, but the checks are inconsistent and spread across a dozen codebases. Compliance is asking for an auditable record of what the agent is and is not permitted to do — something they can review without reading Python. The head of platform security wants a single recommendation before next week's board review — what architectural control would you put in place?

Skills:GovernanceComplianceIdentitySecurity
Attempt 1of 5

Submit sends whatever you have — empty, typed-without-select, or a picked suggestion. Each submission counts as one attempt and unlocks a hint if wrong.