Threat

Cyber hackers use autonomous AI agents to breach over 100 company websites

Page published · Page updated

Dossier

Publication date unknown · First observed: 16 Nov 2026

Coverage timeline

16 Nov 2026beinsure.comobserved

Single-source incident — one report is available.

Why it matters

The reported autonomous AI campaign illustrates how agent-driven attacks can scale website intrusions across many organizations.

Beinsure reports that an unnamed Chinese-speaking cybercriminal used autonomous AI agents to attack more than 1,200 websites and compromise sites belonging to over 100 companies. The report names Deere, SharkNinja and Cummins among the targets.

campaign

Summary

Beinsure, citing a NorForge investigation reviewed by Reuters, reports that a Chinese-speaking operator used autonomous AI agents to attack more than 1,200 websites and compromise sites associated with over 100 companies. The reported operation began around September 25 and targeted e-commerce sites in the United States, Canada, Australia and Japan. Deere, SharkNinja and Cummins were among the companies identified, although their statements described materially different levels of exposure.[0]

The investigation reportedly became possible because operators exposed Cairn multi-agent control panels to the internet. Researchers could inspect task instructions, agent activity and reported outcomes. The largest cluster, designated AIA0001, used Chinese-language instructions and connected 14 panels to more than 1,100 attack projects. Researchers assessed that its principal objective appeared to be persistent access to e-commerce and payment-related systems. The cluster designation does not identify an individual or establish nationality.[0]

The reported activity went beyond a single exploit or routine scanning pattern. Agents investigated alternative infrastructure after unsuccessful initial attempts, adapted exploitation to individual applications and attempted to install upload interfaces associated with web-shell access. Technical examples included a memcached/PHP/Zend Framework chain and cookie-dependent access to a vulnerable backend. NorForge reportedly found a web shell still accessible at one affected supplier, providing technical support for at least some reported successes.[0]

Impact remains bounded by important uncertainties. Deere reported remediation and no indication that sensitive information had been accessed; SharkNinja said the incident affected a standalone support website and had been contained; Cummins was investigating a potential incident involving a third-party vendor's website. The amount of data extracted, if any, was not established. Estimated token costs of approximately $5,000 for the principal campaign—roughly $50 per successful compromise—were researcher estimates, and the apparent AI-model combination was not independently confirmed. All findings in this dossier derive from the single supplied Beinsure article; the underlying investigation and Reuters records were not separately supplied.[0]

Attack chain

  1. 1. Operator tasking: The Chinese-speaking operator reportedly supplied brief instructions through Cairn, including preferred intrusion methods in fewer than 100 Chinese characters. The agents were then left to determine how to achieve the objectives with limited continuing supervision.[0]
  2. 2. Reconnaissance and target expansion: Agents examined target websites and searched for weaknesses. In one documented sequence, unsuccessful testing of a public-facing application led them to examine public technical records, identify additional infrastructure and locate an accessible administrative interface. This was a case-specific sequence, not a demonstrated step at every target.[0]
  3. 3. Iterative, target-specific exploitation: Agents tested alternative exploitation techniques rather than relying on one universal vulnerability. In the administrative-interface case, they reportedly spent hours experimenting before finding a way to execute unauthorized commands. Other documented cases involved a memcached/PHP/Zend Framework chain or a cookie setting that routed requests to a vulnerable backend; these were separate examples.[0]
  4. 4. Attempted persistence and observed residual access: A reported task pattern was to obtain code execution and install a file-upload interface associated with web-shell activity. Researchers assessed persistent access as a principal objective of the largest cluster and separately found a web shell still accessible at a hardware-parts supplier. The report does not establish a common post-compromise sequence or data theft across all affected sites.[0]

Disclosure timeline

DateEvent
Around September 25The reported campaign began directing AI agents against e-commerce websites in the United States, Canada, Australia and Japan. The campaign chronology does not explicitly attach a year to this date.[0]
Around October 2Snorre Fagerland reportedly began identifying exposed Cairn instances used for malicious activity, enabling observation of instructions, agent actions and reported results.[0]
October 19The investigation's findings were reportedly published.[0]
2026-10-26Beinsure published the supplied account, including reported company responses, technical examples and researcher estimates.[0]

How it works

One documented intrusion combined unauthorized write access to a memcached session store, unsafe processing of stored PHP objects and a weakness in older Zend Framework code. The report presents these as components of one application-specific attack chain, not as a universal Cairn exploit. It supplies no CVE, exact affected releases or sufficient implementation detail to map the chain to a particular patch.[0]

At a major US retailer, agents reportedly discovered a cookie setting that directed requests to a vulnerable backend server. This exposed a weakness that was not immediately apparent from the public-facing infrastructure. The article does not identify the cookie, backend product or underlying vulnerability.[0]

In another example, agents moved from an initially unproductive public-facing application to an accessible administrative interface discovered through public technical records. They reportedly identified a subtle processing weakness and iterated until the server executed unauthorized commands. The source does not name the weakness or provide an exact exploit mechanism, so a more specific vulnerability classification is not supported.[0]

Affected versions and patch status

ProductAffectedPatch status
Deere-associated websiteIdentified in the reported unauthorized-access activity; the application, software versions and precise affected scope were not disclosed.Deere said it had taken necessary remediation actions and had no indication that sensitive information was accessed. Specific fixes were not reported.[0]
SharkNinja standalone product-support websiteSharkNinja said the intrusion affected a standalone support-information website rather than its broader corporate infrastructure. Software versions were not disclosed.The company said the threat had been contained; specific patches or remediation details were not reported.[0]
Website operated by a Cummins third-party vendorCummins described a potential cyberattack under investigation, not a confirmed compromise of its broader environment. The vendor, application and versions were not identified.Investigation was reported; no patch or completed-remediation status was supplied.[0]
Application stack using memcached sessions, PHP objects and older Zend Framework codeOne documented intrusion involved unauthorized session-store writes, unsafe PHP-object processing and an older Zend Framework weakness. No exact affected-version range was supplied.No patch identifier, fixed release or remediation status for this particular chain was reported.[0]

Key takeaways

  • The central reported development is sustained, adaptive exploitation with limited operator input—not exploitation of one newly identified vulnerability. The agents reportedly changed approaches and combined target-specific weaknesses over extended tasks.[0]
  • The economic figures require careful qualification. Slightly less than 23 hours per successful intrusion and approximately 10% success were estimates across observed activity involving clusters with different objectives. Roughly $50 per successful compromise was a separate estimate for the largest operator's campaign, not a universal cost benchmark.[0]
  • Chinese-language instructions do not establish nationality or state sponsorship, and an AI worker's name does not confirm its underlying models. The report supports an activity-cluster assessment but not a named-person attribution or definitive DeepSeek/Claude Code configuration.[0]
  • The supplied account provides no actionable attacker infrastructure indicators or exact affected-version-to-patch mapping. Its immediate defensive value lies in the documented exposure patterns, persistence artifacts and distinctions between confirmed access, potential exposure and unestablished data theft.[0]

Defensive actions

  • Scope exposure reviews across e-commerce and payment-related sites, standalone support applications and vendor-operated web properties—not only principal corporate applications.: The largest cluster reportedly sought persistent access to e-commerce and payment systems, while the named-company responses included a standalone support site and a potentially affected vendor-operated website.[0]
  • Hunt on potentially affected web servers for unauthorized file-upload interfaces and residual web shells, and verify that such access has been removed during remediation.: A reported task pattern sought code execution followed by an upload interface, and researchers found a web shell that remained accessible at one supplier.[0]
  • Where the described stack is present, review session-store write permissions, unsafe processing of stored PHP objects and older Zend Framework code as a combined exposure.: One intrusion reportedly chained weaknesses across these components. The source does not provide an exact release-to-patch mapping, so it does not support prescribing a particular fixed version.[0]
  • Assess whether cookie-dependent routing exposes backend servers with different security properties from the default public-facing application, and review administrative interfaces discoverable through public technical records.: Separate reported cases succeeded by reaching a vulnerable backend through a cookie setting and by discovering an administrative interface after initial application testing failed.[0]
  • Determine unauthorized access, persistence and data extraction separately when assessing impact; do not equate a website compromise with confirmed enterprise-wide intrusion or confirmed data theft.: The company statements described different scopes, and the investigation did not establish how much information, if any, was extracted from the principal campaign's targets.[0]