AI agents should have their own identities, tightly limited permissions and no long-lived secrets, says KloudStax CTO Vinay Thakker. His practical advice is to treat an agent less like a clever colleague and more like software that needs carefully fenced access.
Watch Desk analysis
What happened
In an interview with Help Net Security, Thakker argues that every agent should run under its own identity. When it acts for a particular person, its requests should also carry that person’s identity, so the agent can do only what both are permitted to do and the logs record both.
That distinction changes with the job. A support-ticket agent working across a company’s queue should have its own identity, read access to tickets and write access only to routing and labels. An infrastructure agent working for an engineer should carry the engineer’s identity too. It can propose and plan a change, but applying it should still go through the review and deployment process a human would face.
For coding agents, Thakker recommends an allowlist enforced by the system around the model, not a polite request for the model to behave. His proposed starting rules include confining work to a disposable repository, blocking access to credentials, limiting network access to approved destinations, requiring human approval for new packages and consequential changes, and logging every command.
Why it matters
An agent can turn a routine permission into a surprisingly large blast radius. If it reads untrusted material, can access private data and has a way to send information out, a poisoned email or support ticket could put that data at risk, Thakker warns.
His advice also makes clear that a command policy cannot carry the whole security burden. Branch protections belong in version control; network limits in the firewall; credential controls in the identity system. Those checks should still hold if the agent gets around its own rules. A list of instructions is not a security boundary, however sternly formatted.
Our read
The most useful principle here is separation: give the agent its own identity, grant only the access needed for its task, and keep consequential actions behind review. That makes it easier to see what happened and to stop an agent without handing it a person’s entire set of permissions.
Thakker’s recommendations are practitioner advice from an interview, not a guarantee that one checklist makes an agent safe. Still, teams deploying agents can use the details as a concrete starting point, especially the rule against giving agents a person’s login or long-lived secrets. Your AI assistant does not need the keys to the kingdom just because it has found the terminal.
What to watch
- Whether organisations give agents distinct identities and permissions scoped to specific jobs.
- Whether credentials are short-lived and delivered through a controlled proxy rather than exposed to the agent.
- Whether network, repository and deployment controls are enforced outside the agent’s own command policy.
- Whether red-team exercises test quiet data leakage and recovery, not only dramatic overspending or outages.
Discussion spark: Should an AI agent ever be allowed to act with a person’s full permissions, or should every consequential action require a separate approval?
Sources and evidence
- Source update (2 November 2026, 06:30 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.