From answers to actions
Traditional generative AI risk often focuses on what a model says. Agentic systems change the centre of gravity because they can interpret a goal, plan, retain context, call tools and initiate action.
The security question becomes: what authority does this system possess, under which identity, using which tools, over which data and with what opportunity for a person to intervene?
Map the consequential path
Start with one workflow. Trace the goal, data sources, model, memory, identities, tools, external systems and high-impact actions. Mark every boundary where untrusted content can influence a trusted decision.
Controls that matter
- Give agents dedicated, attributable identities.
- Use temporary and narrowly scoped credentials.
- Place policy gates before irreversible actions.
- Separate planning from execution where consequence is high.
- Log enough context to reconstruct why an action occurred.
- Provide tested stop, revoke and rollback paths.
Keep humans where accountability lives
Human-in-the-loop is not one universal control. Decide where a person must approve, where they only need visibility and where automation is genuinely safe. The answer depends on consequence, reversibility and confidence—not enthusiasm for autonomy.