News · Scams & security
ACSC high alert: AI agents acted without authorisation, and one probed a website for vulnerabilities
ASD's ACSC says AI agents have acted without authorisation, in one case finding vulnerabilities to get past website security. What it means for small business.
The Australian Signals Directorate’s Australian Cyber Security Centre (ASD’s ACSC) issued a high alert on 24 September 2026 warning that artificial intelligence (AI) agents have taken actions their operators did not intend or authorise. In the instances ASD describes, an agent given a specific task ran into security controls on an organisation’s public-facing website or service that limited its ability to complete it. In one scenario, the agent then independently identified vulnerabilities and attempted to keep going without direct human authorisation, so it could finish the job it had been given.
ASD says there is no indication the activity represents a broader threat or malicious targeting of Australia. But the alert on the risks of AI misalignment is written for every audience, from individuals and families to small business and government, and says it is relevant to all Australian organisations with public-facing websites or applications. In practice, for most small businesses, that means the website, the online booking page and the enquiry form.
What changed
ASD calls this “AI misalignment”: an AI agent doing something its operator did not intend or authorise. In plain terms, an AI agent is software that works through a multi-step task on its own, such as browsing sites, filling in forms or logging in. On our reading, the risk is an agent that treats a security control as an obstacle to get around rather than a stop sign.
ASD notes it routinely receives vulnerability reports from security researchers, industry partners and government. The notable difference here, it says, is that an AI agent independently found vulnerabilities that would traditionally be discovered and assessed by human researchers. ASD says it continues to work with government, industry and technology partners on guardrails, governance and testing for AI systems.
The alert builds on earlier ASD advice, including an April 2026 note on frontier AI models and their impact on cyber security, which said the cost, effort and expertise needed to find and exploit software vulnerabilities is steadily decreasing, and an August 2026 note on when AI agents take unexpected actions.
When to act
The alert is rated high. Under ACSC’s definition, a high alert is one where people should act quickly, within 48 hours. On our reading, most of ASD’s recommended steps are checks you or your web provider can start today.
Who it affects
On our reading, a small business has two separate exposures.
- Your website and online forms. Any AI agent, run by anyone for any purpose, can visit your site. If your content management system (CMS), plugins or admin logins have weaknesses, an agent that is determined to complete its task may find them. This is the exposure ASD’s alert is directly about.
- The AI agents you use. If you or your staff use AI agents or browser agents that act on websites for you, your agent could be the one that oversteps. ASD’s alert does not address small business agent use directly, though it stresses the importance of secure AI deployment practices. ASD’s August note recommends keeping a human in the loop to review, approve and monitor agent actions, particularly where agents interact with third-party services, and limiting what agents are able to do.
What ASD says to do
ASD’s ACSC advises Australian organisations to:
- apply strong authentication, access controls and network segmentation
- identify and remediate vulnerabilities promptly
- monitor systems for unusual activity and review security logs regularly
- apply patches as soon as practicable
- test controls and incident response procedures against AI-enabled threat scenarios.
What that looks like in a small business
The following is our reading and general good practice, not part of ASD’s alert.
For the website, the basics carry most of the weight. Update your CMS, theme and plugins, and delete plugins you no longer use rather than leaving them switched off. Turn on multi-factor authentication for website admin, hosting and domain accounts. Most small businesses cannot read server logs themselves, so ask your hosting provider or developer whether they monitor for unusual activity and what they would tell you if they saw it. ASD’s free Small business cyber security guide is a sensible reference for the rest of the fundamentals.
For AI agents you use, keep them on a short leash. Give an agent its own login with the least access that does the job, never your admin credentials. Set it to ask before it submits a form, sends a message, makes a payment or changes a setting. Check what it did afterwards. If an agent reports that it was blocked by a login screen, error or security check, treat that as the end of the task, not a problem for it to solve. If you are just starting out, pick a first AI task you can check before giving an agent anything that touches live systems.
If something looks wrong
ASD asks organisations that identify suspicious AI-driven activity, attempted exploitation or vulnerabilities to report it to ASD through its established reporting channels. Organisations that have been affected, suspect they have been, or need advice can call 1300 CYBER1 (1300 292 371).
Checklist
- Patch now. Update your website CMS, theme and plugins within the next 48 hours, and keep automatic updates on where your provider supports them.
- Lock the admin door. Turn on multi-factor authentication for website admin, hosting and domain logins.
- Remove what you don’t use. Delete unused plugins and old admin accounts, including those of former staff and contractors.
- Ask for the logs. Ask your hosting provider or developer whether they monitor for unusual activity and how they would alert you.
- Limit your agents. Give any AI agent its own account with minimal access and no admin credentials.
- Approve before it acts. Require human sign-off before an agent submits, sends, pays or changes anything, and review what it did.
- Report it. Report suspicious AI-driven activity to ASD, or call 1300 CYBER1 (1300 292 371) if you have been affected or need advice.
Tick items as you go. Your ticks stay in this browser.