AI Agents: Security & Controls

Compliance, legal, risk and security-governance teams asked to review AI agents and automation before they go live.

Review an AI agent before it is approved — how it gets attacked, how it fails silently, and which controls actually hold.

Agentic AI has moved from demos into real operational workflows, and the security profile is genuinely different from earlier AI tools. Prompt injection — where instructions hidden in data an agent reads cause it to act against you — has no complete technical fix today, so the control has to sit in what the agent is permitted to do. Reviewers who cannot describe that distinction end up approving exposure they never priced.

• Describe in plain terms what an AI agent can reach, call and change inside your systems
• Recognise the attack paths that matter — prompt injection, data exfiltration and excessive permissions
• Explain why a passing test run is weak evidence that an agent will behave in production
• Specify the permission limits, approval gates and human checks an agent needs before approval
• Challenge a vendor's AI security claims and tell a real assurance from a marketing one
• Run every future agent proposal through a consistent, defensible approval checklist

A starting agenda — every session is shaped around your team, your tools and the risks you’re managing.

1. What an AI agent actually does when it acts on your systems What an agent actually is in engineering terms: the tools it can call, the data it can reach, and the actions it can take without asking anyone.
2. Prompt injection, data exfiltration and the over-permissioned agent The three failure modes that matter most — prompt injection, data exfiltration, and agents given far broader permissions than their job needs.
3. Why you cannot test an agent the way you test software Why traditional testing gives false assurance: the same input can produce different actions, so passing a test run proves less than you think.
4. Designing the leash: permissions, approval gates and human-in-the-loop that hold Designing the constraints that do work — least-privilege permissions, human approval gates on consequential actions, and hard limits on blast radius.
5. Reading a vendor’s AI security claims — and what to ask instead How to read a vendor’s AI security claims, which certifications actually cover agent behaviour, and the questions that get past the marketing.
6. An agent approval checklist your team applies to every deployment A repeatable approval checklist your team applies to every agent, so the review does not depend on who happens to be in the room.

Every team’s needs are different. We’re happy to talk it through and tailor the session to yours — let’s talk →