Malbear Labs said the campaign, designated Operation Snowball, continued for more than two weeks and involved a customised system built around DeepSeek Flash. The researchers reconstructed the activity from an exposed directory containing the attacker’s operational records, rather than from a public disclosure by the affected provider.
The telecommunications company has not been identified. Nor has independent confirmation emerged that all 13 hosted organisations were successfully breached. The number describes organisations identified as targets in the recovered material, an important distinction from verified compromises.
Investigators discovered approximately 70 AI session transcripts alongside reconnaissance findings, stored credentials and operational files. Those records provided an unusually detailed account of how an operator directed an AI system to investigate networks, exploit weaknesses and adjust its actions when attempts failed.
The agent was not acting entirely without human involvement. A human operator established objectives and reviewed results, while the software handled portions of the technical work. The arrangement enabled repeated activity without requiring the operator to issue instructions for every individual command.
Malbear Labs said the system used a modified prompt intended to bypass the model’s normal safeguards. Such manipulation, commonly described as jailbreaking, can induce an AI model to perform tasks that its standard restrictions are designed to discourage.
The investigation identified Cisco BroadWorks XSP as an initial attack surface. Researchers linked the activity to CVE-2026-20320, an XML external entity vulnerability that can allow an unauthenticated attacker to obtain information through improperly processed XML input.
Other vulnerabilities identified in the operation included CVE-2020-7961 affecting Liferay and CVE-2013-2165 associated with JBoss RichFaces. Their appearance in the recovered records illustrates how automated intrusion systems can combine different weaknesses rather than rely on a single entry point.
The attackers also employed established offensive tools, including Sliver for command-and-control operations and tunnelling utilities designed to maintain access across network boundaries. These tools were not themselves evidence of a novel AI-created exploit; the distinguishing feature was the automation coordinating their use.
Researchers said the recovered sessions indicated access to thousands of subscriber billing and device records. However, that assessment depends on material found in the attacker’s workspace. Neither the precise volume of information removed nor the full extent of unauthorised access has been independently established.
Claims involving additional sensitive records also require caution. Without confirmation from the provider, affected customers or an investigating authority, entries in operational transcripts cannot establish that every claimed database was accessed or that its contents were successfully extracted.
The records contained Chinese-language instructions and operational material, but language alone does not establish the nationality, location or institutional affiliation of those responsible. No specific government or organised threat group has been conclusively linked to the campaign.
The disclosure adds to evidence that artificial intelligence can accelerate established cyberattack methods by coordinating reconnaissance, interpreting tool output and retrying unsuccessful steps. It does not demonstrate that the model independently selected the original victim or operated free from human direction.
For telecommunications providers, the reported activity highlights exposure created when customer systems share hosting and management infrastructure. Access obtained through one vulnerable service may give an intruder opportunities to examine other connected environments, although the extent of such movement must be established separately for each organisation.
Security teams investigating comparable activity would need to examine authentication logs, administrative changes, unusual network connections and database access records. Such evidence can distinguish successful intrusions from unsuccessful probes and help determine whether information actually left an affected environment.
The exposed operator directory was an element of the discovery. Rather than observing the attack through alerts, investigators could inspect instructions, outputs and decisions retained by the attacker. That visibility helped reconstruct the relationship between human supervision and automated execution. Nevertheless, records collected from one side of an intrusion may omit failed connections, inaccessible systems or subsequent defensive actions. Their contents therefore require comparison with logs held by the organisations concerned before impact can be determined.
The research was published on 8 November. It did not provide a confirmed campaign start date, identify the affected businesses or document completed remediation by the telecommunications provider. The available material likewise does not establish whether authorities have opened a formal investigation.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
