On this pageWhat AI Changes for OrganizationsFrom AI Assistance to AI AgentsWhere Attackers Use AIImpersonation and Payment FraudDocumented AI-Run IntrusionsThe Time Window in PracticeWhat National Agencies AdviseBusiness AI Agents as TargetsWhat Stays the SameAI Attack Risk Across the BusinessAI Attack Exposure By Organization SizePriorities for DefendersMeasuring Readiness for Faster AttacksAI Attack Defense With ThreatDefenceFrequently Asked Questions Against an organization, it ranges from a language model drafting a phishing message for staff to AI agents working through exposed systems with little human input. The techniques are mostly familiar, and the change for the business is in how little time it has to patch, detect and respond. This guide covers what AI changes for an organization, where attackers use it in an intrusion, documented cases, what national cyber agencies advise, and the priorities for security teams and business leaders. What AI Changes for OrganizationsAI mainly changes the speed, reach and persistence of an attack on an organization. The weaknesses being used are ones security teams already know: exposed services, weak sign-in, unpatched software and excessive access. Speed: The UK’s National Cyber Security Centre (NCSC) assessed in May 2025 that the time between disclosure of a vulnerability and its exploitation had already shrunk to days, and that AI would almost certainly reduce it further. Reach: One operator’s agents can work several organizations at once, so a small or little-known business may receive the same attention as a prominent one. Persistence: An AI agent may repeat and vary its attempts for hours, and typically moves to another endpoint when one starts to limit requests. Access to skill: The NCSC expects AI to raise the capability of novice criminals, hackers for hire and hacktivists, and the Five Eyes agencies say it lowers the barriers for malicious actors. The practical effect is on time. A monthly patch cycle for an internet-facing system, or an alert reviewed the next morning, may leave more time than an automated attacker needs. From AI Assistance to AI AgentsAttackers’ use of AI has moved from help with single tasks to agents that carry a sequence of tasks toward a goal. Public reporting by government agencies and Microsoft marks the stages. January 2024: The NCSC assessed that AI gave attackers an uplift in reconnaissance and social engineering, and judged that more advanced uses were unlikely to be realized before 2025. February 2024: Microsoft reported state-affiliated groups from Russia, North Korea, Iran and China using language models to research targets, draft phishing content and correct scripts, and said it had identified no significant attacks that employed the models it monitored. December 2024: The FBI warned that criminals were using generative AI to commit fraud on a larger scale, with generated text, images, cloned voices and video. May 2025: The NCSC assessed that threat actors were almost certainly already using AI in victim reconnaissance, vulnerability research and exploit development, social engineering, basic malware generation and the processing of stolen data. August 2026: The FBI dates the malicious use of an AI agent framework in real intrusions to at least this month. Each stage took a piece of work away from the operator: first research and writing, then individual technical tasks, then the sequencing of those tasks. The NCSC expected the effect to be a higher volume and impact of intrusions through improvement of existing techniques, with few new ones. Controls that depend on an attacker being slow, selective or short of skill are the ones to review first. Where Attackers Use AIAI appears at most stages of an intrusion, from the first look at a target to the handling of stolen data. The table lists uses that government agencies and Microsoft have described, and what each may look like from inside the organization. Stage How AI is used What the organization may see Reconnaissance Reading an application’s public pages, scripts and API documentation to map its routes and sign-in paths. A burst of requests for script files, source maps and configuration soon after first contact. Social engineering Believable text in any language, cloned voices, and generated images and video used to impersonate a trusted person. Well-written lures, and payment or access requests made by voice or video. Vulnerability research Finding weaknesses and developing working exploits faster. Exploitation attempts soon after a fix or advisory is published. Credential attacks Automated password spraying and abuse of account recovery and sign-up flows. Failed sign-ins across many accounts at a steady rate. Activity after access Testing what a stolen credential can reach, enumerating cloud resources and moving between systems, with each result deciding the next step. Sweeps of discovery calls, many of them denied, and several tools tried against one host. Data theft Searching logs, storage and secret stores for credentials and valuable records. Bulk reads of secrets, logs and files by one identity. Reconnaissance, credential attacks and activity after access are recorded in the organization’s own logs, while vulnerability research and the drafting of a lure show only in their result. Impersonation and Payment FraudGenerative AI makes impersonation cheap and convincing, and the business processes most exposed to it are typically payments, changes to supplier bank details, payroll and requests for access. The FBI’s December 2024 warning describes how each kind of generated content is used. Text: Messages are written faster and reach more people, and translation limits the grammar and spelling errors that once marked a foreign sender. Images: Generated photos support fake profiles, and generated identification documents and credentials support identity fraud and impersonation. Voice: A short cloned audio clip can impersonate a known person, and the FBI notes its use to gain access to bank accounts. Video: Real-time video calls are held with supposed company executives, law enforcement or other authority figures. For a business, the exposure is a payment or access request that appears to come from a senior person. A check that depends on recognizing a voice or a face may pass a forgery, so the control has to sit in the process. Confirm through a second channel. Staff end the call and ring back on a number the organization already holds, which follows the FBI’s advice to hang up and call the organization directly. Separate request from release. A change of supplier bank details or an urgent transfer needs a second approver who took no part in the original conversation. Agree a verification phrase. The FBI suggests a secret word or phrase for families, and the same idea may work for an executive team and its assistants. Apply the rule to resets. A password or multi-factor reset requested by phone gets the same call-back check before the help desk acts. Documented AI-Run IntrusionsThe best documented case of AI agents in real intrusions is ARTEX, a penetration testing framework released as open source in July 2026 for authorized testing. The FBI and Microsoft have both reported on its malicious use against organizations. FBI alert: On November 13, 2026, the FBI issued a FLASH on ARTEX that dates malicious use to at least August 2026 and lists victims in more than ten sectors. State-aligned use: Microsoft tied the framework to Storm-0940, a China state-aligned actor active since 2021, including in a supply chain compromise of a cloud backup service. Criminal use: Microsoft also reported Storm-3358, a financially motivated group it places in Cambodia, using ARTEX against financial organizations. Entry through ordinary flaws: Microsoft reported that Storm-3358 got in through web application weaknesses that the framework and ordinary scanners had located, unauthenticated file upload and server-side request forgery among them. Pace: Microsoft measured 2,900 failed sign-ins a minute from one Storm-0940 address, and cases where the actor moved from first targeting to privileged cloud access in under three hours. Spread: Microsoft has identified threat actors across multiple regions using it, which makes it shared tooling more than the mark of one group. The techniques themselves were established ones. In the Storm-0940 activity the agents automated scanning, password spraying and abuse of OAuth applications. An attack agent of this kind is an AI model connected to tools, memory and a goal, and joint guidance from the Five Eyes agencies describes agentic AI systems as intended to operate without continuous human intervention. Our guide to agentic attack frameworks explains how such a framework is built, what it leaves in sign-in, web and cloud logs, and where the published indicators and hunts are listed. The Time Window in PracticeThe time window is easiest to see by setting an attacker’s clock beside a defender’s. Microsoft measured intrusions in which fewer than three hours separated first targeting from privileged cloud access, and the sequence below places a generic midsize organization inside that window. The times are chosen for the example. 01:00 on a Saturday, first contact. Agents begin reading the organization’s public sign-in pages and web applications. Web and firewall logs record the requests, and nothing yet warrants an alert. 01:20, password spraying. Failed sign-ins arrive across many accounts at a steady rate. The identity provider records each one and may raise a medium-severity alert. 01:50, first success. One password works on an account that was excluded from multi-factor authentication for an old integration. Up to this point the activity was reconnaissance, and from here it is an incident. 02:30, discovery. The signed-in identity lists the cloud resources, applications and secrets it can reach, and many of the calls are denied. 03:40, privileged access. A credential found during discovery gives the agent a privileged cloud role. The outcome depends on what the organization had arranged before 01:00. Next-business-day review: The alert from 01:20 is read on Monday morning, about two days after the agent gained privileged access. Monitoring with approval required: An analyst sees the successful sign-in within minutes and asks for permission to disable the account. Reaching an approver at night may take longer than the agent needs. Monitoring with agreed actions: The analyst revokes the account’s sessions and resets its credential at once, and the run ends before discovery completes. Each of these organizations holds the same logs. The result turns on who is watching at 01:50 and what that person may do without asking. What National Agencies AdviseNational cyber agencies treat AI-enabled attacks as a present risk and ask leaders to act now. Their 2026 publications largely restate established controls with more urgency. Publication Date Issued by What it asks for The AI shift in cyber risk: why leaders must act now June 22, 2026 The cyber security agencies of Australia, Canada, New Zealand, the United Kingdom and the United States A smaller attack surface, faster patching, action on legacy systems, stronger identity and access controls, and incident preparation. Defending against AI-enabled cyber attacks, guidance for government, critical infrastructure and large enterprises July 6, 2026 ASD’s Australian Cyber Security Centre Recommended actions grouped by urgency, most of them for internet-facing services. Careful adoption of agentic AI services May 1, 2026 ASD’s Australian Cyber Security Centre with CISA, NSA and the Canadian, New Zealand and UK cyber centers Incremental deployment of AI agents on low-risk tasks, with strict privilege controls, continuous monitoring, strong identity management and human oversight. The Five Eyes statement is direct about time: “The timeline is not years, it is months.” Its actions are established practice, and the statement singles out operational systems with long update cycles and unsupported systems, which it calls easy targets. The statement also says leaders must be confident that controls will perform during a real incident, that this requires reassessing long-standing trade-offs, and that unsupported systems are strategic liabilities as well as technical debt. The NCSC’s May 2025 assessment adds a view of how the threat may develop in the period to 2027. Known vulnerabilities: AI-enabled tools will almost certainly improve attackers’ ability to exploit known vulnerabilities, which raises the volume of attacks on systems that lack current security fixes, and the NCSC names operational technology with lower levels of security as particularly exposed. A digital divide: The NCSC expects a divide between systems that keep pace with AI-enabled threats and a large proportion that are more vulnerable. Critical systems: Assuming a lag or no change in mitigations, it sees a realistic possibility that critical systems become more vulnerable to advanced actors by 2027. The Australian Signals Directorate (ASD) turns the same themes into specific actions, such as naming an accountable executive and identifying internet-facing services daily. Our guide to the ASD guidance on AI-enabled cyber attacks covers each action in the document’s order. The agencies also expect defenders to use AI. The Five Eyes statement says adversaries are already using it and that defenders must do the same, and it lists what AI in security operations can deliver: earlier detection of vulnerabilities, better software quality, monitoring for unusual behavior and faster incident response. Business AI Agents as TargetsAn organization’s own AI agents add to its attack surface. The joint guidance on agentic AI notes that agents inherit the weaknesses of the language models they are built on, and that every connected tool, data source and memory store widens the exposure. Prompt injection: Instructions hidden in content the agent reads, such as a phishing email processed by an email-monitoring agent, may direct it to take a harmful action. Excess privilege: An attacker who compromises a tool in an agent’s workflow can inherit the agent’s privileges, including any that exceed what its task requires. Stolen agent identity: Agents authenticate with keys and tokens, and actions taken under a trusted agent’s identity may evade audit controls and detection. The agencies recommend that organizations avoid granting agents broad or unrestricted access, especially to sensitive data or critical systems, and use them for low-risk tasks. They also advise managing AI security inside existing cyber security frameworks. The guidance illustrates the privilege risk with a procurement agent that was given broad access to financial systems, email and contract repositories. An attacker who compromised one low-risk tool in its workflow could then modify contracts and approve payments, and the audit logs appeared legitimate because the actions ran under the agent’s identity. The same document sets out practices for deploying and operating agents, and the steps below draw on them. Start narrow. Deployment is phased, with access and autonomy increased progressively and rolled back after failures. Keep approval with people. Designers or operators decide when human approval is required, and the agent cannot carry out high-impact actions such as system resets or deletion of critical records without it. Use short-lived credentials. Static, long-lived secrets are replaced with credentials that expire when the job is complete, and privileges are held to the minimum the task requires. Isolate the agent. Segmentation limits the effect of a failure, and the agent has no write access to logs. Monitor the whole operation. Operators log tool calls, decisions and actions as well as inputs and outputs, and audit identity and privilege changes for drift. Plan for compromise. Incident response procedures cover detecting, containing and recovering from a compromised agent. The NCSC makes a related point in its May 2025 assessment: prompt injection, software vulnerabilities and supply chain attacks can already be used against AI systems to reach the wider systems they connect to. What Stays the SameThe entry points, the evidence and the effective controls stay largely the same. AI-powered attacks succeed through weaknesses that already exist, and they are recorded in the same logs as any other intrusion. Entry points: Weak passwords, exposed application files, open self-registration, stolen application credentials and long-lived access keys remain the usual ways in. Controls: The Five Eyes agencies state that success comes from getting the basics right and acting quickly, with secure-by-design practice and defense in depth. Human direction: The NCSC judged in May 2025 that fully automated, end-to-end advanced attacks were unlikely in the period to 2027, with skilled operators staying involved. Agent errors: Agents repeat themselves, try many tools where one would do, and leave test artifacts behind, which gives a security team more to detect. AI Attack Risk Across the BusinessAI-enabled attacks concern more of the organization than the security team, because the decisions that shorten exposure sit with several functions. The Five Eyes agencies call the risk a core business risk and a leadership responsibility. Function Decision it owns Why it matters Board and executives Who is accountable for the risk, and whether cyber leaders have the authority and resources to act. The Five Eyes statement asks leaders to understand and assess risk, readiness and accountability. IT operations How quickly internet-facing systems are patched, and when unsupported systems are retired. Patch windows are usually set by availability targets that IT operations owns. Finance and executive support Which payment and access requests need confirmation through a second, known channel. A cloned voice or generated video can pass a check that relies on recognizing the person. Security operations Which containment actions may be taken without waiting for approval. An approval chain adds delay at the point where minutes count. Supplier management Which suppliers hold access to systems, and how that access is reviewed. A supply chain compromise was among the documented cases. AI Attack Exposure By Organization SizeThe exposure is similar at every size, and the means of reducing it differ. Automation removes the protection a low profile once gave, so each organization needs an answer to who watches its internet-facing systems and how quickly they are fixed. Small business: The exposed systems are typically a website, email and a few cloud services, often maintained by an outside provider. ASD’s guidance for small businesses asks the owner to turn off websites that are no longer needed, turn on automatic security updates, and name a person or service provider who looks after the security of those websites. Midsize organization: A small IT team typically carries security alongside other work, so monitoring outside business hours and patch speed are the usual gaps. ASD’s guidance for medium-sized businesses starts by asking for the document to be discussed with the CEO, the security and IT staff, and the organization’s cloud and managed service providers. Large enterprise: The difficulty is breadth. Many business units publish internet-facing services, legacy systems are slow to change, and a complete inventory depends on continuous external attack surface management. MSP or MSSP: A provider is a target in its own right and a route to its clients. In the supply chain compromise Microsoft reported, the actor attempted sign-ins to more than 2,250 Microsoft 365 tenants. Priorities for DefendersThe priorities below are in order, starting with what an automated attacker reaches first. Each follows the agency guidance described above. Know what is exposed. Keep a current inventory of internet-facing services and remove those that are unnecessary. Patch exposed systems faster. Shorten the patch cycle for internet-facing systems, and replace or isolate products that have lost vendor support. Strengthen sign-in. Require phishing-resistant multi-factor authentication for privileged accounts and review permissions on a schedule. Limit what one credential can reach. Give service accounts and applications the minimum access they need, and rotate long-lived keys. Detect behavior and respond quickly. Collect sign-in, cloud, web and endpoint logs centrally, and write detections on behavior, since source addresses and user agents change quickly. Test as often as systems change. Continuous penetration testing checks new and changed systems between scheduled tests, so the organization finds an exposed weakness first. Rehearse the response. Exercise incident response and recovery plans, and measure how quickly an intrusion can be contained. An organization that finds ARTEX activity is asked by the FBI to report it to a local field office or to IC3. Measuring Readiness for Faster AttacksReadiness for a faster attacker can be measured with times and coverage figures taken from records the organization already keeps. Measure What it shows Where the record is Inventory lag How long a new internet-facing service exists before the security team knows of it. DNS and cloud change logs compared with the first entry in the asset inventory. Time to patch exposed systems Days from the vendor’s fix to deployment, for each internet-facing service. Vendor advisory date compared with change records and patch tool reports. Multi-factor exclusions Accounts and applications that can still sign in with a password alone. Identity provider authentication method report and the exclusion list of each access policy. Payment verification The share of bank detail changes and urgent transfers that carry a recorded call-back. Finance system audit trail and the approval record for each change. Agent permissions What each of the organization’s own AI agents can reach, and the age of the credentials it uses. Service account and application records in the identity provider, and the agent platform’s configuration. Authority outside business hours Which containment actions the person on call may take without approval. Response runbook and the on-call roster. The figures are more useful when they are cut the way an automated attacker works. Split by hour: Nights, weekends and holidays are reported separately, since an agent works at the same pace at 03:00 as at 15:00. Report the slowest case: An average hides the one system that took a month to patch, and that system may be the one an automated scan finds. Test the measure: An authorized test or exercise that raises a real alert shows the actual response, where a policy document shows the intended one. Detection coverage and the time taken to contain an agent run are measured separately, and the guide to agentic attack frameworks covers those measures and the response steps. AI Attack Defense With ThreatDefenceThreatDefence puts the attacker’s method to work for the defender. Offensive agents test a customer’s exposed systems continuously, while a 24x7 SOC watches the customer’s logs for the same kind of activity. Both run on one AI SOC platform that covers security monitoring, compliance and cyber risk management. Exposure: The ASM agent maps what the organization exposes to the internet and records each new or changed host. Testing: The testing agent works inside a scope the customer has authorized and keeps proof of each weakness it shows to be exploitable. Earlier use: For monitored customers, the Defender and Hunter agents check telemetry for signs that the weakness was used before it was found. Oversight: SOC analysts review significant findings, higher-impact tests wait for approval, and every request and result is logged. A test starts when something changes, such as a new advisory for an exposed product or a leaked credential, and on a rolling schedule. Agentic penetration testing describes the service. Frequently Asked QuestionsWhat is AI hacking?AI hacking is the use of AI to prepare or carry out a cyber attack against an organization. It covers generative AI that writes phishing messages or clones a voice, AI that helps find and exploit vulnerabilities, and AI agents that run tools against a target and decide each next step. The term is also used for attacks on AI systems themselves, such as prompt injection.Can AI carry out a cyber attack without a human?Partly. In the documented agent-run intrusions an operator still chose the target and the goal, and the agents carried out most of the steps in between. The UK’s NCSC judged in May 2025 that fully automated, end-to-end advanced attacks were unlikely in the period to 2027, with skilled operators remaining in the loop.Do AI cyber attacks use new techniques?No. The documented AI-run intrusions relied on established techniques, including password spraying, abuse of OAuth applications and exploitation of exposed web applications. AI changes how fast those techniques are applied and how many targets one operator can work at once, which is why agencies ask for faster patching and less exposure ahead of new categories of control.Are smaller organizations targets of AI-powered attacks?Yes. Automation lets one operator work many targets at once, so the size or profile of an organization offers little protection. The NCSC expects the spread of AI-enabled tools to widen access to intrusion capability, and ASD notes that AI can be used to target many systems at once, especially internet-facing services such as websites.Can AI-generated voice or video be used to defraud a business?Yes. The FBI warned in December 2024 that criminals use generative AI to clone voices and to generate video for real-time calls with supposed company executives and other authority figures. A payment or access request made by voice or video is safer when staff confirm it through a second channel the organization already trusts, such as a known phone number.What do government agencies recommend against AI-enabled attacks?The Five Eyes cyber agencies ask leaders to reduce the attack surface, accelerate patching, address legacy systems, strengthen identity and access controls, and prepare for incidents before they happen. ASD’s guidance for Australian organizations turns those themes into ordered actions, most of them for internet-facing services. Both treat the work as established practice that has become urgent.Can attackers hack an organization's own AI agents?Yes. The joint Five Eyes guidance on agentic AI describes prompt injection, where instructions hidden in an email or web page direct an agent to take a harmful action, and compromise of the tools and credentials an agent uses. The agencies advise limiting agents to low-risk tasks, restricting their privileges and keeping human oversight.How fast can an AI-run attack move?Fast enough that a next-day review may come too late. Microsoft measured 2,900 failed sign-ins a minute from one address in the Storm-0940 activity, and intrusions that reached privileged cloud access less than three hours after first targeting. The NCSC assessed in May 2025 that the time from disclosure of a vulnerability to its exploitation had already shrunk to days.When did attackers start using AI?Public reporting dates from early 2024. The NCSC assessed in January 2024 that AI gave attackers an uplift in reconnaissance and social engineering, and Microsoft reported in February 2024 that state-affiliated groups were using language models for research, phishing content and scripting. The FBI dates malicious use of an AI agent framework in real intrusions to at least August 2026.Is an organization that uses no AI still exposed to AI hacking?Yes. AI hacking describes the attacker’s tooling, and it is aimed at ordinary systems: sign-in pages, web applications, cloud accounts and staff inboxes. An organization with no AI products of its own faces the same faster scanning, password spraying and phishing. Adopting AI agents adds a second exposure, since those agents can be attacked too.