ASD Guidance on AI-Enabled Cyber Attacks Explained
Skip to content

ASD Guidance on AI-Enabled Cyber Attacks

The ASD guidance on AI-enabled cyber attacks is a short publication from the Australian Signals Directorate’s Australian Cyber Security Centre (ASD’s ACSC), first published on July 6, 2026.

On this page

It lists recommended actions for large organizations, grouped by urgency, and most of them concern internet-facing services.

This guide explains who the publication is for, what it says about AI in attacks, each recommended action in the document’s order, the versions for smaller organizations, the Essential Eight comparison, and a practical first month.

Who the Guidance Is for

The guidance is written to help government, critical infrastructure providers and large enterprises defend against cyber attacks that involve the use of artificial intelligence. Its full title is Defending against AI-enabled cyber attacks: Guidance for government, critical infrastructure and large enterprises.

The publication page names the readers inside those organizations: executives, boards, cyber security leaders and technical teams.

ASD published a version for medium-sized businesses and a version for small businesses on the same day, and a later section of this guide compares them.

How AI Changes Attacks

ASD describes the change in a short opening statement about speed, scale and chaining, and says malicious actors may use AI in the following ways.

  • Discovery and exploit development: AI may be used to discover vulnerabilities and to develop exploits within hours.
  • Many systems at once: Attackers may target many systems in parallel, and ASD singles out internet-facing services such as websites.
  • Chained vulnerabilities: A chain of lower severity vulnerabilities may be exploited to cause a high severity impact.
  • Rapid progression: An attack may move quickly from initial compromise to data theft, extortion or disruption.

Each point removes a margin that defenders have relied on. A monthly patch cycle assumes weeks between disclosure and exploitation, and triage by severity assumes that a medium-rated finding can wait. A low public profile assumes that an attacker has to choose targets.

The document stops at that summary. It names no threat groups, techniques or incidents, and the guide to AI hacking covers how attackers use AI and AI agents, with the documented cases.

How the Actions Are Organized

The recommended actions sit in groups named by urgency: immediate, short-term, medium-term and long-term. The document gives no dates or durations for the groups, so each organization sets its own deadlines.

  • Scope: Most actions name internet-facing services, and one medium-term action extends the same treatment to other devices by risk assessment.
  • Strength of wording: Immediate and short-term actions are written as direct instructions, while several medium-term actions begin with “Consider” and the long-term group holds a single action on architecture.

The document gives the first action to a person and others to a meeting, and the rest divide in practice among the readers the publication page names.

  • Cyber security leader: The security function owns the vulnerability management process, verification of controls, log analysis and the plans.
  • Technical teams: The teams that run each service carry out daily identification, least privilege, segmentation and configuration baselines.
  • Procurement and development teams: The supplier actions change how software is bought, and the code review action changes how it is built.

Immediate and Short-Term Actions

The immediate and short-term actions reduce what an attacker can reach and prepare the organization to recover. The table follows the document’s order and summarizes the ASD text. The last column, here and in the later table, holds suggestions from practice.

Group Action What it asks Where to start
Immediate Named executive Appoint a named executive accountable for the risks from malicious actors using AI and from the organization’s own use of AI. Record the appointment and the reporting line to the board.
Immediate Appetite for brief outages At an executive committee or board meeting, accept more risk of internet-facing services being briefly unavailable through rapid remediation, more frequent patch windows, or a bad patch. Name each service and the outage it can tolerate.
Immediate Rapid patching Update vulnerability management to rapidly patch or mitigate vulnerabilities of all severities in internet-facing services, automatically where the risk of a bad patch is acceptably low or it rolls back easily. List the services that can take automatic updates today.
Immediate Appetite for unsupported services At the same level, reduce the appetite for compromise of internet-facing services that lack vendor support, and treat their removal or replacement as far more urgent. List unsupported services with an owner and a date.
Immediate Daily identification Identify all internet-facing services daily, remove the unnecessary ones, progress removing or replacing unsupported ones, and add enhanced monitoring and compensating controls meanwhile. Automate discovery from DNS, certificates and cloud accounts.
Immediate People and resources At the same level, allocate suitable and sufficient personnel and resources to address the risks from AI, including these actions. Cost the actions and present the gap.
Short-term Least privilege and segmentation Give accounts used by internet-facing services the minimum access and privileges, and segment each service from the others and from the corporate network. Begin with service accounts that hold administrator rights.
Short-term Secure configuration baselines Develop, deploy, enforce and monitor baselines that disable or remove unnecessary functionality, insecure settings, and unsupported or weak communication protocols. Write the baseline for the most common platform first.
Short-term Verification of controls Verify that patching, segmentation, least privilege accounts, configuration baselines and other technical controls on internet-facing services work as implemented. Test from the internet as well as reviewing configuration.
Short-term Central logging and automated response Centrally capture and promptly analyze security logs from internet-facing services, with automated triage and automated response actions that are pre-approved, narrowly scoped and reversible. Confirm every service sends logs, then agree the approved actions.
Short-term Incident response plans Prepare and exercise plans to remediate incidents rapidly, including restoring from backups that are recent, tested and protected from unauthorized access, modification and deletion. Restore one service from backup and time it.
Short-term Continuity and recovery plans Prepare and exercise business continuity and disaster recovery plans, including testing the ability to rapidly block network traffic and reset all user passwords. Rehearse a full password reset on paper, then in a test.

The daily cadence is the most demanding instruction in the immediate group. Discovery has to be automated to meet it, and the guide to external attack surface management explains how that discovery works.

Decisions for the Board

Half of the immediate actions are decisions for an executive committee or board of directors meeting. They change what the organization is willing to accept, which a security team cannot do alone.

The risk appetite decisions move in opposite directions, with more tolerance for a short outage caused by fast patching and less for compromise through unsupported services.

A board paper that covers these decisions may include the following.

  • Services and outage tolerance: The paper lists the current internet-facing services, marks the unsupported ones and states the downtime each can absorb.
  • Patching rule: It proposes a time from the release of a fix to its deployment, and the conditions for automatic patching.
  • Removal plan: It gives each unsupported service a date and an owner, with the compensating controls in place until then.
  • Resources: It states the people and budget the actions need.

The appointment of the named executive covers more than attacks. The same person is accountable for the risks from the organization’s own use of AI, so a board may ask for a register of where AI is used alongside the list of internet-facing services.

Rapid Patching in Practice

The patching action changes more daily work than any other in the document, since it covers every severity and prefers automatic deployment. ASD describes the outcome and leaves the procedure to each organization.

The sequence below is one practical reading. It follows a medium-severity vulnerability in a self-hosted file transfer service through the actions.

  1. The service is already listed. The daily identification run records it with its version, owner and vendor support status.
  2. The vendor releases a fix. The patch is rated medium, so a rule based on severity would hold it for the monthly window. The ASD wording puts it in scope at once.
  3. The patching rule sets the route. The fix deploys automatically where a brief outage is accepted for this service and rollback is easy. Otherwise it waits for the next patch window behind a mitigation, such as disabling the affected feature.
  4. The patch is applied and checked. A health check confirms the service after the update, and a failed check starts the rollback.
  5. The fix is verified from outside. A scan or a test from the internet confirms that the weakness is closed, which is the short-term action on verifying controls.
  6. Logs are searched for earlier use. The service’s security logs are searched for requests matching the vulnerability from before the fix, since the document says exploits may be developed within hours. The change record then holds the release and deployment times and both results.

The medium rating matters in this example. ASD warns that lower severity vulnerabilities may be chained, so a rule that hurries only critical fixes leaves the links of a chain in place.

Medium-Term and Long-Term Actions

The medium-term actions concern where software comes from and how AI is used in defense, and the long-term action concerns architecture.

Group Action What it asks First step
Medium-term Software as a service Consider software-as-a-service cloud services for internet-facing services, from secure and transparent providers that are responsible for rapidly remediating vulnerabilities on the organization’s behalf. Identify self-hosted services with a mature hosted equivalent.
Medium-term Supplier commitment Use software from suppliers with a demonstrated commitment to security and transparency, so that fewer vulnerabilities need remediation and patches or mitigations arrive promptly. Add patch history and disclosure practice to procurement criteria.
Medium-term Separate security patches Use software from suppliers that securely provide adequately tested patches separate from optional feature updates, so automatic patching carries a low risk of supply chain attacks and faulty or incompatible patches. Ask key suppliers whether security fixes ship apart from features.
Medium-term AI for code review Consider using AI, in a secure, controlled and human-supervised manner, to find and confirm vulnerabilities in code built for internet-facing services, before release and periodically in production. Pilot on one application, with a developer confirming each finding.
Medium-term AI for logs and testing Consider using AI, under the same conditions, to assist with analyzing security logs from internet-facing services and with penetration tests and vulnerability assessments of them. Define scope, limits and human review before any AI-run test.
Medium-term Other computing devices Consider, based on a risk assessment, which other devices would benefit from timely patching, segmentation, least privilege accounts, configuration baselines and prompt log analysis. Rank internal systems by what an attacker would reach next.
Long-term Modern defensible architecture Progress to a modern defensible architecture aligned to a layered architecture, with clear traceability of designs to business, security and resilience objectives, using zero trust principles and Secure by Design practices. Read ASD’s modern defensible architecture publications and set a roadmap.

The supplier actions support the patching rule in the immediate group. Automatic patching is safest when security fixes arrive tested and separate from feature changes.

AI in Defense and Testing

ASD treats AI as a tool for defenders as well as attackers, under stated conditions. Both medium-term actions on defensive AI carry the same qualifier, and the guide to AI penetration testing explains how AI-run tests work and the controls that keep a person in charge.

The short-term group points the same way without naming AI, in its call for automated triage and response. Blocking one address or ending one session would typically meet the conditions it sets for an automated action.

The July document leaves the words secure, controlled and human-supervised undefined. ASD’s later publication on AI in cyber defense, listed under related guidance below, was published on September 8, 2026 and co-sealed by the cyber security centers of Canada, New Zealand and the United Kingdom. It describes the practices behind those words, including the following.

  • Human oversight: A person approves actions with significant security, operational or safety impact, and autonomous actions are limited to those that are narrowly scoped, preapproved and reversible.
  • System protection and sandboxing: AI tools run with minimal privileges inside sandboxes and execution limits, and their actions are traceable and reviewable.
  • Testing and assurance: A tool is tested in the organization’s own environment before production use, since vendor claims and demonstrations are insufficient evidence.
  • Data readiness: Asset inventories, logging coverage, identity data, vulnerability records and configuration data are assessed and improved before and during adoption, since AI-supported defense is only as reliable as the data it draws on.

The limit on autonomous actions repeats the conditions that the July guidance sets for automated response. The later publication also describes a red team using an AI model to link low and medium severity weaknesses into attack paths, the defender’s use of chaining.

It cautions that identifying vulnerabilities does not, on its own, strengthen security, which also takes context, prioritization and remediation. Its appendices list questions for vendors of AI-enabled security tools, covering measurable outcomes, auditability, data sovereignty and behavior when the AI component fails.

Versions for Smaller Organizations

The versions for medium-sized businesses and for small businesses address the same threat with fewer and simpler actions. The table summarizes how each version treats the same subject.

Subject Large organizations Medium-sized businesses Small businesses
Accountability A named executive, with decisions taken at an executive committee or board meeting. A discussion of the guidance with the CEO, cyber security and IT personnel, and service providers. The publication is shared within the business, and someone is made responsible for the security of its websites.
Exposed services All internet-facing services identified daily, and unnecessary ones removed. Internet-facing services identified regularly, with website servers and email servers as examples, and unnecessary ones removed. Websites that are no longer needed are turned off, and unneeded features are disabled on the rest.
Patching All severities patched or mitigated rapidly, automatically where a bad patch is unlikely or easy to roll back. All security patches applied rapidly, automatically if the risk of a faulty patch is low. The website maintainer is asked whether patches are automatic or applied rapidly after release.
Staff devices Considered in the medium term with other computing devices, by risk assessment. Web browsers and operating systems on user devices apply security patches automatically. Automatic security updates are turned on for computers, tablets and mobile phones.
Logs Captured centrally and analyzed promptly, with automated triage and response. Security logs from internet-facing services reviewed promptly. The website maintainer is asked how quickly a compromise would be noticed and reported.
AI for defenders Offered as something to consider, for code review, log analysis and testing. Written as an instruction to use AI carefully, for the same purposes. Absent. A business that develops software is asked to scan it for vulnerabilities.

The response and continuity actions are almost identical in the large and medium versions. Other differences run against expectation. The medium version makes automatic patching of user devices an immediate action and states the AI actions more firmly, while the long-term architecture action appears only in the large version.

A medium-sized organization with its own security team may treat the medium version as the minimum and the large version as the target.

Working With Service Providers

Each version of the guidance gives part of the work to a provider. The large version makes a software-as-a-service provider responsible for rapid remediation, the medium version opens with a discussion that includes service providers, and the small version asks the business to identify who is responsible for the security of its websites.

The small version then lists questions for that maintainer.

  • Vendor support: The maintainer confirms that vendors still provide security patches for the website’s software, and whether unsupported software can be replaced.
  • Patch speed: The maintainer confirms that all security patches are applied automatically or rapidly after they become available.
  • Software and configuration: The maintainer confirms that the software comes from a reputable company and has unnecessary functionality and insecure settings disabled.
  • Detection and recovery: The maintainer states how quickly a compromise would be noticed and reported, and whether it can fix the issue, evict the intruder and restore from a recent secure backup.

The same version asks owners to review the companies that store and share their customer data, and points to ASD’s Questions to ask managed service providers. That publication asks whether a provider implements the Essential Eight, administers its systems securely, monitors activity, assesses its systems regularly and is prepared to respond to incidents.

An MSP or hosting company can expect these questions from clients of every size. A written answer would typically hold the client’s list of internet-facing services, the patching rule with its evidence, and the time taken to notice and report a compromise.

The Guidance Beside the Essential Eight

The July guidance makes no reference to the Essential Eight, although many Australian organizations already measure themselves against it. The comparison below sets the guidance beside Maturity Level One of ASD’s Essential Eight maturity model, which uses the term online services.

  • Patch timeframe: The maturity model allows 48 hours for a vulnerability the vendor assesses as critical or one with a working exploit, and two weeks for the rest. The guidance asks for rapid patching or mitigation at every severity and gives no number.
  • Discovery cadence: The maturity model asks for automated asset discovery at least fortnightly and a vulnerability scan of online services at least daily. The guidance asks for all internet-facing services to be identified daily.
  • Unsupported services: The maturity model states that online services no longer supported by vendors are removed. The guidance makes any that remain a matter for the board, with compensating controls until they are gone.

An organization that already meets those timeframes has most of the process the guidance needs, and the main change may be to move vulnerabilities rated below critical onto the faster path.

ASD’s Information Security Manual sets its own intervals for vulnerability assessments and penetration tests, which the guide to penetration testing frequency lists with the other standards.

Evidence for Each Action

Each action leaves a record that an executive or an auditor can inspect. ASD prescribes none of them, so the table is a practical suggestion.

Action Record Where it is kept
Executive and board decisions The minute naming the executive, the risk appetite decisions and the resources allocated. Board minutes and the risk register.
Rapid patching For each service, the date a fix was released and the date it was applied. Patch management tool and change records.
Daily identification Dated inventory snapshots and the daily difference. Attack surface management tool or discovery job output.
Unsupported services The end-of-support date, owner, removal date and compensating controls for each. Asset register.
Least privilege and segmentation Each service account’s permissions, and the rules between services and toward the corporate network. Identity platform, firewall and cloud configuration exports.
Verification of controls Dated test results that show each control working. Penetration test and vulnerability assessment reports.
Logging and automated response The log source list with last event times, the approved automated actions and a record of each run. SIEM and response tool audit logs.
Response and recovery plans Exercise records, a timed restore test, and the results of a test network block and full password reset. Exercise reports and the backup platform.

The records support a few measures the named executive can report. These are suggestions as well, since the document sets no metrics.

  • Patch time: The time from the release of a fix to its deployment on each internet-facing service, reported for every severity.
  • Unsupported services: The number still reachable from the internet, and the oldest removal date among them.
  • Recovery time: The measured time to restore a service from backup, to block network traffic and to reset all user passwords.

A Practical First Month

A first month can cover the immediate group and schedule the short-term work. The sequence below is one reasonable order.

  1. Week 1, ownership and inventory. The executive is named in writing. The security team lists internet-facing services from DNS records, certificates, cloud accounts and firewall rules, and marks each as needed, unnecessary or unsupported.
  2. Week 2, removal and monitoring. Unnecessary services are switched off. Each unsupported service gets an owner, a removal date, and added monitoring and compensating controls until it is gone.
  3. Week 3, board decisions. The paper goes to the executive committee or board. The patching rule changes once the outage tolerance is agreed, with automatic patching where rollback is easy.
  4. Week 4, daily operation and dates. Discovery runs every day and differences are reviewed. The short-term actions each receive an owner and a date, starting with a restore test and a rehearsal of a full password reset.

Acting on the Guidance With ThreatDefence

For Australian customers of ThreatDefence, data sovereignty and residency in Australia are guaranteed, and the SOC team is based in Australia. That is relevant where a measure involves handing logs, inventories or test evidence to a provider. Several measures map to work done on the ThreatDefence AI SOC platform, which brings security monitoring, compliance and cyber risk management together, and by its 24x7 SOC.

  • Daily identification: The ASM agent maps internet-facing domains, hosts, services and certificates, and logs each change as it appears.
  • Verification of controls: The testing agent tests exposed systems within an authorized scope, keeps proof of the weaknesses it confirms, and retests when a fix is reported.
  • Human-supervised AI: Higher-impact tests wait for approval, every request and result is recorded, and SOC analysts review significant findings.
  • Prompt log analysis: The Defender and Hunter agents search a monitored customer’s telemetry for signs that a confirmed weakness was used earlier.

Agentic penetration testing describes the testing service in full.

Frequently Asked Questions

Is the ASD guidance on AI-enabled cyber attacks mandatory?

No. The publication describes its contents as guidance and recommended actions, and it refers to no legislation, standard or compliance deadline. An organization with obligations under other frameworks or laws still has to meet those separately. The document may be read as a statement of what ASD wants large organizations to prioritize.

When did ASD publish the guidance?

ASD’s ACSC first published it on July 6, 2026, and the download on the publication page is labeled July 2026. Versions for medium-sized businesses and for small businesses carry the same first published date. The current files on cyber.gov.au are the ones to rely on, since ASD updates its publications over time.

Does the guidance apply to small and medium-sized businesses?

No. This version is addressed to government, critical infrastructure providers and large enterprises. ASD published separate versions for medium-sized businesses and for small businesses on the same day, with fewer actions and simpler wording. The actions described in this guide come from the version for large organizations.

What does ASD mean by internet-facing services?

The document does not define the term. It gives websites as its example, and the version for medium-sized businesses adds website servers and email servers. In practice the term covers any service that accepts connections from the internet, which typically includes remote access gateways, exposed management interfaces, APIs and cloud-hosted applications.

How fast does ASD expect vulnerabilities to be patched?

The document sets no number of hours or days. It asks organizations to rapidly patch or mitigate vulnerabilities of all severities in internet-facing services, automatically where the risk of a bad patch is acceptably low or the change can be rolled back easily. It also says attackers may develop exploits within hours, which indicates the pace ASD has in mind.

How long are the immediate, short-term and medium-term periods?

The document does not say. It orders the actions by urgency and leaves the dates to each organization. A reasonable reading is that immediate actions begin at once, since most of them are decisions and process changes, while medium-term actions such as changing suppliers depend on contract and procurement cycles.

Does ASD recommend using AI for cyber defense?

Yes, as something to consider. The medium-term actions suggest AI for finding and confirming vulnerabilities in the organization’s own code, for assisting with security log analysis, and for assisting with penetration tests and vulnerability assessments of internet-facing services. Each mention carries the condition that the use is secure, controlled and human-supervised.

Why does the guidance ask boards to change risk appetite?

Fast patching occasionally breaks a service, and many organizations delay patches to avoid that. ASD therefore places the trade between availability and security with the executive committee or board, where the business impact of an outage is owned. Once the decision is minuted, the team has a mandate to patch quickly and to retire services that have lost vendor support.

Does the ASD guidance mention the Essential Eight?

No. The publication for large organizations names no other framework or standard. Its long-term action refers to modern defensible architecture, zero trust principles and Secure by Design practices. Its patching and discovery actions go further than Maturity Level One of the Essential Eight maturity model in places, for example by asking for rapid patching at all severities.

Is there a version of the ASD guidance for managed service providers?

No. ASD published versions for large organizations, medium-sized businesses and small businesses. Providers appear in each as parties to consult or to question, and the small business version lists questions for whoever maintains a business’s websites. A provider can read those questions as the standard its clients are told to expect.

Talk to Our Experts

See how ThreatDefence automates security monitoring, compliance and cyber risk management in one platform.

    I'm an MSP, MSSP, or reseller

    By submitting this form, you accept the terms in our privacy policy.