Palo Alto: 40% of firms can't retrace what their AI agents did

By DataForge team17 November 20265 min read

Palo Alto Networks found 40% of organisations can't reconstruct an AI agent's actions; here's how NZ firms can fix agent identity and logging.

What the report found

Palo Alto Networks published its 2026 State of Cloud Security Report on 16 November 2026. It's the sixth edition, and it surveyed 2,800 security leaders and practitioners across 10 countries and 13 industries. Most of the report covers AI agents in software delivery. Palo Alto says 99% of organisations rely on agents for autonomous or semiautonomous actions in their development pipelines, and 73% let agents modify code.

The accountability figures matter most for owners and IT leaders. According to Palo Alto, 40% of organisations can't fully reconstruct an AI agent's actions after it has run, and 53% log agent activity under a human developer's identity or personal access token. SiliconANGLE's coverage reports that the share unable to reconstruct agent actions rises to 46% in financial services. The same coverage says 60% of organisations still need manual intervention to revoke access for non-human identities. Revoking or rotating a compromised developer credential, API key or pipeline secret takes an average of 2.5 days.

Palo Alto's post doesn't name the 10 countries, so nobody outside the company knows whether New Zealand or Australian respondents are in the sample. Palo Alto also sells products aimed at these gaps, and readers should weigh the findings with that in mind. Techzine's report on the same release gives matching identity and logging figures. The underlying problem is simple enough to check in your own environment without relying on anyone's survey.

Why borrowed identities hurt an investigation

When an agent runs under a developer's token, the audit log records the developer as the actor. Techzine notes that investigators then have to separate human actions from agent actions before they can work out what happened during an incident. In a small team, that might mean interviewing staff to find out whether a change came from them or from a tool they'd connected weeks earlier.

The delay matters in New Zealand because of the Privacy Act 2020. The Office of the Privacy Commissioner says organisations must notify the Commissioner when a privacy breach has caused, or may cause, serious harm. Its guidance sets an ideal of notifying within 72 hours of becoming aware of a notifiable breach. Suppose customer records left a database and the log only shows a developer's account. The investigating team then has to work out whether a person or an agent pulled the data before it can judge who obtained the information and whether the organisation still controls it. Both questions feed into the serious harm assessment, and a muddled audit trail slows both.

A borrowed identity also widens what an agent can reach. A developer's personal token usually carries all of that developer's permissions. An agent that only needs to open pull requests in one repository inherits access to everything else the developer can touch. The report found that only 9% of organisations can automatically discover and classify ungoverned AI across their cloud environments, which suggests many firms don't have a complete list of the agents holding those inherited permissions.

What this means for NZ businesses

The survey focuses on development pipelines, but the same pattern turns up wherever staff connect an agent to a business system using their own login. Coding assistants, workflow builders, service desk tools and finance platforms increasingly let an agent act inside the system on a user's behalf. In each case, the system's log shows the person who granted access as the actor.

Most NZ organisations don't have a security operations centre that can untangle those logs after the fact. They'll get better value from fixing identity at setup than from buying more detection. Palo Alto's own recommendations come down to ordinary identity hygiene: give each agent a distinct identity, enforce least privilege, monitor agent activity, and use short-lived credentials with automated revocation. Most organisations can do much of that with the identity provider and admin consoles they already pay for, though agent and service accounts may carry their own licence costs, so check before rolling them out widely.

Agent identity also belongs in change and access management. If an agent can change code or configuration, its access should go through the same approval and review steps as a new staff member or contractor. Firms that run their service desk on an ITIL 4 based process can add agent accounts to their existing access request and periodic review workflows without inventing a separate process.

Australia's direction adds weight for trans-Tasman firms. On 10 November the Australian Government announced planned AI safety and security protections, including mandatory reporting of AI-related cybersecurity incidents among the review recommendations it accepted. A firm that can't say what its agents did will struggle to meet any incident reporting rule, whichever side of the Tasman the rule comes from.

Practical next steps

  • Build an agent register. List every AI tool that can act in a business system, the human account or token it uses, and the systems that account can reach. Start with code repositories, cloud consoles, Microsoft 365 and your finance platform.
  • Move agents onto their own identities. Create a dedicated service account or workload identity for each agent, name a human owner for it, and stop agents running on personal access tokens.
  • Cut permissions to the task. Scope each agent identity to the repositories, mailboxes or ledgers it needs. Use read-only access wherever the agent doesn't need to write.
  • Switch to short-lived credentials. Set expiry dates on agent tokens and API keys, and write down who can revoke each one out of hours. Test whether you can revoke an agent's access in minutes, then compare your result with the 2.5-day average in the report.
  • Check that logs separate agents from people. Run an agent task and confirm that your audit log records the agent identity as the actor. Keep logs long enough to support a privacy breach assessment.
  • Add agents to your breach response plan. Update your incident runbook so the first triage step asks whether an agent identity was involved, and practise against the Privacy Commissioner's 72-hour guidance.

If you're planning wider agent use and want the identity and logging design settled before rollout, our AI agents team can help set up the register and access model.

Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.

Get my free assessment How we help
Get new articles by RSS