Excessive Agency in Agentic AI: Setting Safe Boundaries Project | Quiz

WHY: New agents should start highly restricted because it is far easier to grant additional permissions than to recover from excessive agency incidents. CONTEXT: As agents demonstrate reliability over time they can earn expanded permissions with full audit trail. Any misuse should trigger automatic permission downgrade. REMEMBER: Start restrictive expand carefully.

WHY: The agent chains multiple legitimate capabilities together for malicious purposes becoming both the vulnerability scanner and the attack vector in one autonomous sequence. CONTEXT: Broad tool access allows capability chaining where each tool is legitimate individually but devastating in combination. REMEMBER: Chained capabilities equals the agent becomes scanner exploiter and exfiltrator.

WHY: This is a Tool Chain Cascade where multiple connected tools allow a single error to propagate through the entire system. CONTEXT: The agent access to multiple connected systems means one mistake chains through commit build and deploy before humans can intervene. REMEMBER: Connected tools equals cascading risk.

WHY: This is Persistent Behavioral Drift where an agent behavior gradually changes over time without explicit authorization. CONTEXT: The danger is incremental: not a sudden change but slow expansion that goes unmonitored. Eventually high-risk operations are executed autonomously. REMEMBER: Gradual boundary expansion equals behavioral drift.

WHY: The five layers are: Least Privilege Tool Restrictions Human-in-the-Loop Behavioral Monitoring and Clear Boundaries. CONTEXT: No single control is sufficient you need multiple overlapping layers for defense in depth. All five layers working together provide comprehensive protection. REMEMBER: All five layers not just one.

WHY: Agents can be manipulated through prompt injection can misinterpret well-crafted prompts and can find unexpected paths to goals. CONTEXT: Prompt engineering safety training and good intentions are not security controls. Real security comes from external constraints like permissions monitoring and approval workflows. REMEMBER: You cannot prompt-engineer your way to security.

WHY: Agentic AI does not just advise it acts. It calls APIs modifies databases sends communications and executes transactions. CONTEXT: Traditional AI predicts and suggests while humans execute. With agentic AI mistakes cause real-world damage before humans can intervene. REMEMBER: The gap between AI suggests and AI does is where excessive agency becomes dangerous.

WHY: OWASP states Treat agent requests like requests from the internet because agents might be compromised misinterpret instructions or pursue goals through unexpected means. CONTEXT: This means never assume agent behavior is safe verify every action independently and enforce security at system boundaries not in agent logic. REMEMBER: The agent is not on your team do not trust verify.