Safety for AI Agentss
Agent safety is the set of controls that stop an agent from being manipulated or from causing harm through its tools. The main threats are prompt injection, excessive permissions and data leakage, so defences focus on least privilege, input and output checks, isolation and human approval.
Stage: Quality and Safety · Part of the AI Agents roadmap · Last reviewed 19 November 2026
What to learn in Safety
- Explain how indirect prompt injection reaches an agent through retrieved content
- Apply least privilege to tool credentials and data access
- Separate untrusted content from instructions in the prompt
- Require confirmation for actions that send, delete or pay
- Log tool calls so abuse can be audited later
Safety interview questions
What is indirect prompt injection?
It is an attack where malicious instructions are hidden in content the agent reads, such as a web page, email or document, rather than typed by the user. The model may obey them. Defences include isolating untrusted text, limiting tool powers and confirming risky actions.
Why is least privilege important for agent tools?
An agent can be tricked or can simply err, so any permission it holds may be misused. Giving each tool read-only or narrowly scoped credentials limits the damage of a bad call. It is easier to add permissions later than to recover from a breach.
Where Safety fits in the AI Agents roadmap
Safety belongs to the Quality and Safety stage (advanced level). Trust.
Other units in this stage
Back to the full AI Agents roadmap