AI agents can inherit broad permissions, act across connected systems and keep running after their original use has ended. Forbes Technology Council contributor Nidhi Jain recommends four practical controls: give each agent its own task-scoped identity, require human approval for high-risk actions, log what agents do across systems, and set an expiry date for every agent.
Watch Desk analysis
What happened
Jain, founder of AI governance company CloudEagle.ai, argues that security controls need not mean blocking AI adoption. Her rule of thumb is to match friction to risk: reading data can proceed without a checkpoint, while actions such as moving money, sending external messages or changing production systems should require approval before execution.
For organisations already using agents, she also suggests taking an inventory across their platforms, checking OAuth grants against agent identities, classifying agents by the actions they can take, and retiring agents without a named human owner. These are recommendations in a Forbes Technology Council essay, not requirements imposed by a regulator.
Why it matters
The useful distinction is between what an agent can see and what it can do. A prompt log may record what someone asked; action-level logs can show what the agent actually changed across connected services. That makes it easier to investigate problems and to decide whether an agent’s permissions still fit its job.
Expiry dates tackle a quieter problem: access can outlive the person or project that created it. A six-month review gives teams a chance to renew useful agents with fresh checks and remove the rest, rather than leaving old credentials to gather dust with the keys still in them.
Our read
This is a refreshingly workable checklist, not a demand to put every agent behind a human tollbooth. Start with a permissions inventory and identify which agents can write, send, spend or change production systems. Read-only agents and agents with authority to alter the business should not be treated as the same security problem.
What to watch
- Whether organisations make agent-level identities and action logging defaults rather than deployment chores.
- Whether high-risk actions receive approval before execution, not merely an alert afterwards.
- Whether scheduled reviews lead to real access changes and expired agents being decommissioned.
Discussion spark: Should AI agents that can change records, send messages or spend money always need approval before acting, or can carefully scoped permissions make that unnecessary?
Sources and evidence
- How CISOs Can Set Guardrails On AI Agents Without Killing Adoption – Forbes (3 November 2026, 14:30 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.