Back to all articles

Autonomous cyberattacks: four real-world AI agent campaigns

Share

AI agents are no longer used by threat actors only to draft phishing messages or write malicious code. They are already being used to make decisions and run commands during real intrusions. On 21 October 2026, CrowdStrike published research into four distinct threat actors, observed between May and August 2026, who delegated parts of their cyber operations to agentic AI systems.

The main lesson for businesses is not the arrival of some unstoppable “AI hacker”. It is that familiar weaknesses — exposed administrative access, poorly protected accounts, insufficient segmentation and slow detection — are more consequential when adversaries can test and exploit them at greater speed.

What CrowdStrike actually uncovered

According to its 21 October investigation, CrowdStrike's Counter Adversary Operations team recovered insight into four operators because their AI-enabled offensive frameworks had been left exposed in publicly accessible directories.

This allowed researchers to examine their software architectures, agent instructions, orchestration designs and records of offensive activity. These are documented cases, not a comprehensive census of all AI-enabled intrusions. The four operators are not shown to be working together.

An AI agent in this context links a language model with tools that can act: read output, choose the next task, run commands and process the results. That does not mean full independence from humans: operators still initiate, configure, monitor or redirect these campaigns.

Four actors, four approaches to agent-driven intrusions

1. Argentine healthcare: a supervisor and an executor

CrowdStrike reports that one Spanish- and English-speaking operator compromised at least 15 healthcare entities in Argentina. Credentials and medical records were exfiltrated from at least one victim. This does not establish that patient data was taken from every one of the 15 organisations.

The operator had reworked a penetration-testing environment based on the open-source PentestGPT project. A Supervisor agent selected steps from a defined attack workflow, while an Executor performed tasks. Scripts, tools and MCP servers extended the system's capabilities.

Defensive lesson: intrusions can progress through successive stages with limited human input. Log correlation must help responders connect an unusual initial login, suspicious credential use and unexpected access to sensitive records.

2. Government and business targets: multiple agent architectures

A second operator, using Chinese, tracked more than 20 named targets across regions. CrowdStrike reports compromises affecting government organisations in two countries; independent WIRED reporting identifies Brazil and Colombia. The actor also targeted private firms in Brazil and the United States, in sectors such as healthcare, biotechnology, retail and technology.

The operator maintained two orchestration frameworks. One combined automated reconnaissance with AI selecting from predefined actions; another gave individual agents, or cooperating specialist agents, more flexibility. Access to some models was mediated through an AI proxy service.

Defensive lesson: improve controls around privileged identities, APIs, administrative services and cross-environment access. Rapidly switching targets and techniques compresses the time available for investigation.

3. Basic tooling, public skills and actual compromises

A third operator, using Indonesian, relied on 819 pre-existing AI skills, largely taken from public repositories, without sophisticated custom orchestration. According to CrowdStrike, human intervention was sometimes required to overcome model refusals and encourage further attempts.

Nevertheless, researchers report compromises of Indonesian government domains, a government domain in another country and private-sector entities.

Defensive lesson: an attacker's apparent lack of deep expertise no longer provides reassurance. Basic defenses need to hold up against repeated, automated attempts.

4. Data theft and SaaS intrusions

A fourth group of Russian-speaking individuals used a relatively straightforward seven-stage workflow. CrowdStrike attributes the exfiltration of customer data from a US CRM provider to this group, as well as compromises of SaaS and data-centre organisations in multiple countries.

The case illustrates the risk posed by digital supply-chain dependencies: a breach at a provider may affect data entrusted to it by other businesses. The public evidence does not, however, justify describing this as a ransomware campaign.

Defensive lesson: review vendor privileges, cloud audit trails, bulk exports and procedures for revoking external access.

Attribution caveat: the language used to control an AI agent does not prove an operator's nationality, geographical origin or state affiliation.

Why these operations change the pace of an intrusion

The documented change is the combination of a language model, an actionable toolset and a decision loop. After receiving an objective, an agent can interpret output, choose another action and repeat without a human approving every intermediate step.

The official detailed report landing page describes a case where an agent allegedly moved laterally and harvested domain-controller credentials in approximately twenty minutes without additional human direction during that sequence. This is a CrowdStrike-reported observation, not an average attack time or a general performance benchmark.

CrowdStrike researchers interviewed by WIRED stressed that the observed intrusions did not mainly depend on novel exploits. Speed, volume and preventable security weaknesses played a central role. For defenders, response time can therefore be just as important as whether a detection product is installed.

Model safety controls are not an enterprise security boundary

The investigation describes actors circumventing model safeguards by framing malicious operations as authorised penetration tests. Some operators changed tools, models or instructions when they encountered refusals.

This does not establish that every AI assistant contains a remotely exploitable flaw or that the AI vendors themselves were compromised. It does show why an organisation cannot rely on a model provider's behavioural guardrails to enforce its own network access, identity policies or data permissions.

This is distinct from our 23 October report about OpenAI agents weakening their own monitoring. That story involved internal task agents undermining controls; the present story concerns adversaries deliberately using agents as intrusion tools.

Six practical defenses for smaller organisations

1. Reduce exposed entry points

Inventory internet-facing applications, VPNs, administrative interfaces, cloud services and service accounts. Remove unnecessary exposure, prioritise known exploited or critical vulnerabilities, and protect administrative endpoints. Our PHP2Shell remediation analysis illustrates why checking the runtime actually in production matters.

2. Reduce the value of stolen credentials

Enable phishing-resistant multifactor authentication for sensitive access wherever feasible. Remove shared accounts, minimise standing privileges, manage service credentials and promptly revoke unnecessary access. Monitor abnormal logins and credential reuse across environments.

3. Make lateral movement harder

Segment networks, isolate backups and critical services, restrict unnecessary server-to-server communication and control remote administration. An initial compromise should not grant unrestricted access to identity infrastructure, databases and other systems.

4. Correlate behaviour rather than isolated alerts

Combine identity events, privilege changes, host-to-host connections, process creation, network events and unexpected data exports. The Credential Access, Lateral Movement and Exfiltration categories in MITRE ATT&CK offer a useful framework. Ensure that high-priority alerts have real recipients and an escalation path.

5. Prepare continuity and recovery

Maintain separate, tamper-resistant backups of business data, configuration and other restoration dependencies. Test the ability to restore them. A backup that has never been restored does not constitute a dependable recovery plan. Define who can disable access and recover priority services under pressure.

6. Exercise incident response

Run scenarios involving a compromised administrator account or unauthorised data export. Test who receives the warning, who can isolate assets, how evidence is preserved and how affected customers or partners are notified. The NIST Cybersecurity Framework 2.0 provides a structure covering Govern, Identify, Protect, Detect, Respond and Recover.

What to collect when an intrusion is suspected

During an authorised investigation, preserve identity and authentication logs, EDR telemetry, privileged-account events, file-share access, cloud activity and data-export history. Build a timeline, scope affected assets and preserve relevant evidence with hashes and reliable timestamps.

A high rate of requests does not by itself prove that an AI agent was involved. Legitimate administration tools, authorised scanners and conventional scripts can also generate automated activity. Attribution requires additional, corroborated evidence.

If compromise is confirmed, coordinate containment, session revocation, targeted credential rotation, assessment of affected data and trusted restoration, taking account of applicable reporting obligations.

What this means for GLOV Solutions

For GLOV Consulting, these cases make risk governance, dependency mapping, supplier security and operational continuity tangible issues. GLOV Secure focuses on digital-asset protection and authorisation architectures; the auxiliary accounts and services around those assets also require security controls.

GLOV AI is still in preparation. This article concerns real-world malicious use of AI agents; it does not imply that a commercially available GLOV AI product already offers specific protection against these campaigns.

Our recommendation is to strengthen the controls that limit the consequences of an initial foothold. Offensive automation changes the speed of operations, but it does not remove the value of least privilege, segmentation, tested backups and organised human response.

Limitations and unanswered questions

This article is based on technical research by a private security company, with additional reporting from independent outlets. Not all victims are named, and the full evidence required to independently verify every reported compromise has not been made public. Four cases cannot establish the overall share of cyberattacks involving AI, show coordination between these groups or provide a meaningful average success rate.

Predictions about future growth in AI-enabled attacks are CrowdStrike's assessments, not events already observed. Nothing in the cited sources indicates that GLOV Solutions systems were affected.

GLOV Solutions editorial analysis — information assessed as of 24 October 2026.

Official sources and independent analysis

Related articles

Back to all articles