A city CISO requires threat scenarios for agent goal hijacking before a production 311 agent rollout. What is the CISO demanding?
Select an answer to reveal the explanation.
Short Explanation
Before the 311 agent goes live, the CISO wants war stories on paper — how someone could rewrite its goals or instructions. That is threat awareness for goal hijacking, not optimism that prompts always win. Name the scenarios before residents meet the bot.
Full Explanation
A city CISO requiring threat scenarios for agent goal hijacking before a production 311 agent rollout is demanding explicit threat awareness and scenarios covering goal and instruction hijacking for civic agents. Goal hijacking and instruction manipulation are first-class risks for agentic systems that plan and call tools; production launch without those scenarios leaves 311 refunds, case reopen, and resident messaging exposed to unexamined abuse paths. The CISO is insisting on documented threat modeling, not blind trust in policy prompts.
Skipping threat analysis because agents always follow municipal policy prompts fails by concept: prompt compliance is not guaranteed under injection, tool-description attacks, or conflicting instructions. Disabling all logging so hijack attempts leave no embarrassing evidence removes detection and forensics rather than addressing hijacking threats. Granting the 311 agent unrestricted tool scopes to reduce friction at launch increases impact if goals are hijacked.
Exam caveat: threat scenarios must be specific to agent goal/instruction hijacking (poisoned tools, prompt injection, supervisor-prompt conflict, rogue planner steps)—generic malware-catalog checklists alone are insufficient for GAIPS-style agentic controls. Operational check: before go-live, review a written threat matrix with at least several hijack scenarios, mapped mitigations (runtime allowlists, human approval for high-impact tools, kill switch), and a tabletop that exercises one goal-hijack path against the 311 agent design.