Agentic AI

Why it matters

• An agent's mistake is an action, not a sentence: a payment, a deletion, an email or a code change.
• Where an agent is part of a high-risk system, EU AI Act duties on human oversight and logging govern how much authority it is given.
• Every tool and permission is attack surface, so an agent registry and a least-privilege review belong in the release process.

In detail

Agents combine a model with tools, memory and permissions. The main security risk is excessive agency (OWASP LLM06): an agent with more functionality, permissions or autonomy than its task needs can cause damage when its output is wrong, ambiguous or manipulated, for example through prompt injection. Controls include least-privilege tool access, authorisation enforced outside the model, rate limits, logging, and human approval for consequential actions.

Laws and policies on record that use it

• Singapore starter kit for testing agentic AISingapore · Guidance
• Singapore agentic AI governance frameworkSingapore · Voluntary standard

Related concepts

• Prompt injectionPrompt injection is an attack in which crafted input, typed by a user or hidden in content the model reads, ov...
• Human oversightDesign and operating measures that let people effectively oversee an AI system while it is in use: understand...
• OWASP Top 10 for LLM ApplicationsThe OWASP Top 10 for LLM Applications is the Open Worldwide Application Security Project's list of the ten mos...