AI agents are no longer just answering questions. In many organisations they now write code, change configurations and push updates into live systems. A new survey from Palo Alto Networks suggests most businesses have handed over those keys faster than they have put controls around them.
What the report found
Palo Alto Networks' State of Cloud Security Report 2026 surveyed 2,800 security leaders and practitioners in 10 countries. According to SiliconANGLE and Help Net Security, the headline figures are:
- Agents are everywhere. 99% of organisations use AI agents for autonomous or semi-autonomous actions in their development pipelines.
- They can change things. 73% allow agents to modify code, pipeline configurations or production environments.
- Nobody can see them all. 99% found ungoverned AI in their environment in the past year, but only 9% can automatically discover and classify it.
- Their actions are hard to trace. Two in five organisations cannot reconstruct what an AI agent did, and 53% say agent activity is logged under a developer's own identity or personal access token.
- Cutting off access is slow. 60% still need manual steps to revoke access for non-human identities, and it takes an average of 2.5 days to revoke or rotate compromised credentials, API keys or pipeline secrets.
- Data is leaking into the wrong tools. 40% said governed data had been copied into ungoverned AI tools.
Confidence is high all the same: 97% said they were at least somewhat confident they could secure their AI supply chain, yet only 36% can monitor AI behaviour in production.
Why this matters beyond software teams
The survey focuses on development and cloud teams, but the pattern applies more widely. AI assistants that read mailboxes, AI tools connected to CRM or finance systems, and automation platforms that act on behalf of staff all raise the same question: what can this tool change, and who would notice if it changed the wrong thing?
When an agent works under a person's login, its mistakes look like that person's mistakes. When it holds a long-lived API key, losing that key can mean losing control of whatever the agent can reach. And if nobody owns the tool, nobody is watching it.
Six steps to take control
- List every AI agent and assistant. Include tools your teams adopted on their own. For each one, record what it connects to and what it can read, write or change.
- Give each agent its own identity. Do not let agents act under a staff member's account or personal token. A separate identity makes their actions visible and easy to switch off.
- Grant only the access each agent needs. Read-only by default. Changes to production, finance or customer records should require a human approval step.
- Log what agents do. Make sure agent actions are recorded in a way your team can search and review, so you can reconstruct events if something goes wrong.
- Name an owner for every tool. Someone should be accountable for each agent's permissions, its updates and the data it touches.
- Be ready to cut access quickly. Know how to revoke an agent's credentials in minutes, not days, and rotate API keys and secrets on a schedule.
Our view
AI agents can save real time, and few businesses will want to give them up. But the same rule applies to them as to any new member of staff: clear identity, limited access, someone responsible and a record of what was done. Putting that in place now is far easier than untangling it after an incident.
If you would like help reviewing the identities, access and logging around the AI tools in your business, contact our team.