AI agents can turn an ordinary-looking instruction into a real-world change if they have the permissions to act. Huawei’s practical advice for security teams is to inventory those agents, limit their access and require human approval for high-impact actions.
Huawei AI Watch analysis
What happened
A Huawei-authored guide published by GovInsider recommends mapping each agent’s owner, purpose, connected systems, data access, permissions and external tools. It says to prioritise agents that can take actions, rather than only make recommendations.
The guide gives a concrete example: an agent processing supplier invoices could be misled by hidden text in an email into changing bank details. Its proposed controls include narrowly scoped, expiring permissions; approval for elevated access; and human checks before consequential or irreversible actions. It also recommends keeping tamper-resistant logs of what agents access and do.
What to do now
- Inventory agents and their access
Record each agent’s owner, purpose, connected systems, permissions and tools, including those embedded in other software. - Limit data and permissions
Give agents only the access needed for a defined task, for only as long as it is needed. - Require approval for consequential actions
Treat changes such as updating payment details or deleting records differently from routine tasks. - Make actions reconstructable
Log the systems, tools and identities an agent used, what it did, and any approvals or overrides. - Include personal devices in the rules
Restrict sensitive workflows where a device cannot meet security requirements, and make access revocable if it is lost or compromised.
Why it matters
The advice focuses on the agent’s permissions and surrounding workflow, not just whether the underlying model is secure. That matters because an agent using legitimate access can still make a damaging change if it follows a malicious instruction or has more authority than its task requires.
Huawei also points to visibility across agent activity as a key challenge and mentions its AgentArts tools. The guide is practical security advice from a company with products in this space, not an independent evaluation of those tools.
Our read
The strongest takeaway is simple: an agent that can recommend a payment change is not the same as one that can make it. Start by listing agents that can act, then check what they can reach and which actions need a person’s approval. It is not glamorous work, but neither is explaining an avoidable transfer.
What to watch
- Whether organisations publish clear inventories and owners for agents already in use.
- Whether sensitive or irreversible actions consistently require human approval.
- Whether monitoring can show, after an incident, what an agent accessed and changed.
Discussion spark: Should organisations require human approval for every high-impact action an AI agent can take, even if that slows routine work?
Sources and evidence
- How CISOs can secure agentic AI environments without slowing innovation – GovInsider (3 November 2026, 00:31 UTC)
not affiliated with or endorsed by Huawei AI