Australia's Cyber Agency Says Go Slow on AI Agents — What It Means for Your Business

TL;DR: The ASD's cyber centre has co-signed official guidance telling organisations to use AI agents only for low-risk, non-sensitive tasks for now. It lands just as agentic tools go mainstream — and it's sensible advice for any Australian business, not only government.

What changed

In early May 2026 (30 April US time), the Australian Signals Directorate's Australian Cyber Security Centre (ASD's ACSC) joined the United States' CISA and NSA, plus the cyber authorities of Canada, New Zealand and the United Kingdom, to publish a joint guide titled "Careful Adoption of Agentic Artificial Intelligence (AI) Services." It is the Five Eyes' first joint guidance specifically on agentic AI — the systems that don't just answer a question but take actions on your behalf.

The headline recommendation is restraint. Organisations should start with lower-risk, non-sensitive tasks and widen an agent's autonomy only as security controls mature. The guide also presses for built-in security controls, least-privilege access, and strong identity and access management for AI agents — treating an agent that can click, log in and transact as a system to be governed, not a chatbot to be trusted.

Why it matters for Australians

This is no longer abstract. Agentic tools have gone mainstream in 2026: Perplexity's Comet browser ships free on every platform, Genspark places real phone calls on your behalf, and Zoom's AI Companion now runs workflows. The capability is exactly what the guidance is cautious about. Our own testing found the same risks the ASD flags — Comet's agent can be hijacked by hidden instructions on a web page, and Genspark's autonomous actions don't always succeed. An agent with access to your inbox, bank login or CRM is a far bigger exposure than a tool that only drafts text.

For a small Australian business, the practical takeaway is simple: let AI agents handle low-stakes, reversible work, keep a human in the loop for anything sensitive, and never hand an agent standing access to a logged-in financial or customer account without sign-off.

Does this change our recommendation?

No — it sharpens it. We already score agentic tools cautiously on reliability and security, and this guidance is the official version of the advice we've been giving in every agentic review: use them, but supervise them. See our Comet vs Genspark comparison for where each one is safe to lean on, and where it isn't.

Sources

1. ASD's ACSC / Cyber.gov.au — "Careful adoption of agentic AI services" — cyber.gov.au — Wayback: [snapshot queued] — checked 26 June 2026 — Wayback 9 June 2026
2. Joint guide PDF (US DoD mirror) — media.defense.gov — checked 26 June 2026 — Wayback 10 June 2026
3. NSA press release — nsa.gov — checked 26 June 2026

Read our Comet and Genspark reviews for how far to trust each agentic tool →