SecurityWeek’s weekly cybersecurity news roundup offers a concise, overarching view of vital developments that may not always receive standalone coverage, yet remain profoundly relevant to the broader threat landscape. This curated summary highlights critical stories spanning vulnerability disclosures, emerging attack methods, policy updates, industry reports, and notable law enforcement actions to help readers stay well-informed about the rapidly evolving global cybersecurity environment.
PoeLLM Malware Finds Its C&C Server Hidden in a GitHub Poem
Among the most inventive threats detailed this week, threat intelligence researchers at Black Lotus Labs have uncovered an unusual malware variant dubbed PoeLLM. Active since at least April 2026, PoeLLM specifically targets exposed artificial intelligence and open-source services—primarily focusing on LiteLLM, Ollama, Gotenberg, and Gitea instances—with the dual objectives of mining cryptocurrency and expanding a burgeoning botnet.
What makes PoeLLM particularly novel is its command-and-control (C&C) infrastructure. Infected machines pull a specific poem hosted on GitHub, extract four designated keywords from the text, and programmatically convert them into the IP address of the active C&C server. This architecture allows the operators to dynamically shift their command infrastructure simply by altering the words in the repository’s poem. Researchers note that the operator, assessed with moderate confidence to be an Italian speaker, has already updated the underlying GitHub poem eleven times as part of ongoing operational adjustments.
GhostAction Secret Theft Spreads to 772 More GitHub Repositories
Supply chain security continues to present severe structural risks to software development ecosystems. According to monitoring data from GitGuardian, the GhostAction supply chain campaign aggressively expanded its footprint between August 31 and September 30. During this window, the threat actors pushed their secret-stealing GitHub Actions workflow into 772 additional public repositories, affecting 373 distinct users and organizations.
The campaign systematically targeted 2,577 sensitive secrets, including hardcoded SSH keys alongside Azure, AWS, and database credentials. Apart from adopting a new exfiltration server to route the stolen data, the attackers largely reused their established 2025 playbook. GitGuardian’s telemetry indicates that this malicious campaign has operated persistently without significant operational downtime.
Jury Convicts Uranium Finance Hacker in Multi-Million Dollar Crypto Heist
In a major milestone for federal cybercrime enforcement, a federal jury has convicted 36-year-old Jonathan Spalletta of Maryland on multiple counts of computer fraud and money laundering. The conviction stems from two devastating 2021 cyberattacks targeting Uranium Finance, a decentralized cryptocurrency exchange. Spalletta successfully exploited smart contract flaws to siphon approximately $1.4 million in the initial intrusion, followed by a staggering $53.3 million in a subsequent attack that ultimately forced the exchange to shut its doors permanently.
Following the thefts, Spalletta laundered the vast array of stolen digital assets through complex cryptocurrency transactions involving privacy-enhancing tools like Tornado Cash. He then converted portions of the illicit proceeds into high-end physical collectibles, including rare Magic: The Gathering and Pokémon cards, as well as antique Roman coins. He now faces up to 10 years in prison on the computer fraud count and an additional 20 years on the money laundering count when he is sentenced.
CISA Narrows Cybersecurity Retention Pay Criteria
Shifting toward federal policy and workforce management, the Cybersecurity and Infrastructure Security Agency (CISA) has announced plans to maintain its Cybersecurity Retention Incentive program through fiscal year 2027. However, the agency is implementing significantly stricter qualifying criteria. Moving forward, participating staff must maintain an "exceeds expectations" performance rating or higher, and they must dedicate at least 51% of their working hours to core cyber duties within one of three designated job series.
Employees outside those specific series who spend 75% or more of their time on cyber work must now apply through an official review board for consideration. This structural overhaul comes directly in the wake of a Department of Homeland Security Inspector General finding, which concluded that the previous iteration of the retention incentive program was mismanaged and applied far too broadly across the agency.
Coalition Maps Out Federal OT Requirements for CISA Directive
Addressing critical infrastructure vulnerabilities, the Operational Technology Cybersecurity Coalition (OTCC) has published a detailed proposal calling for a binding operational directive (BOD) from CISA specifically focused on operational technology within federal civilian agencies. These agencies collectively rely on more than 8,000 General Services Administration-managed facilities housing essential heating, ventilation, and air conditioning (HVAC), power management, physical access control, and building automation systems.
The OTCC proposal urges CISA to require federal agencies to designate a senior official or specialized office directly accountable for OT security compliance. Furthermore, the directive would compel agencies to apply existing security requirements while prioritizing CISA’s established Cybersecurity Performance Goals to protect vital government facilities from physical and digital disruption.
Domino’s Resets Accounts Hijacked via Recycled Passwords
In the consumer retail sector, fast-food giant Domino’s began notifying a subset of customers that their online accounts had been accessed by unauthorized third parties. The breach was executed via credential stuffing, a method where attackers test email and password pairs leaked in unrelated, third-party data breaches against active accounts on the platform.
Domino’s emphasized that its internal corporate systems were never compromised during the incidents and that the platform does not store sensitive customer payment details. As a precautionary measure, the company automatically reset the passwords on all affected user accounts to prevent further unauthorized access.
AI Under Suspicion as South Korea Probes Bank Cyberattacks
Artificial intelligence is facing intense scrutiny following a wave of sophisticated cyberattacks targeting South Korea’s financial institutions. South Korean President Lee Jae Myung publicly stated that investigators have found strong preliminary indications that AI technologies were actively utilized in some of the recent intrusions, which resulted in the exposure of sensitive customer personal information.
National police have launched a comprehensive, full-scale investigation into the breaches. While authorities have not yet publicly disclosed the specific AI tools deployed or the complete financial scope of the data loss, private cybersecurity researchers are shedding light on the mechanics. CrowdStrike reported uncovering Claude Code session histories, ARTEX configuration files, and Claude memory files while examining the threat actor’s operational infrastructure. Based on these findings, the security firm assesses with moderate confidence that a Chinese-speaking, financially motivated threat actor group is behind the campaign.
Empire Market Co-Founder Sentenced to Four Decades Behind Bars
In landmark dark web enforcement news, 30-year-old Raheim Hamilton of Virginia has been sentenced to 40 years in federal prison and ordered to pay a $5 million fine. Hamilton previously pleaded guilty to a drug conspiracy charge tied to his role as co-creator and operator of Empire Market, a notorious dark web marketplace that operated between 2018 and 2020 alongside co-founder Thomas Pavey.
During its operational lifespan, Empire Market facilitated over four million illegal transactions valued in excess of $430 million. While the platform primarily functioned as an illicit marketplace for narcotics sales, it also served as a clearinghouse for stolen credentials, personal identification information, counterfeit currency, and specialized hacking tools. Pavey, who entered a guilty plea last year, is scheduled to receive his sentencing later this month.
Exposed Nvidia DCGM Exporters Leak Telemetry From 12,000 GPUs
Vulnerability disclosures this week highlighted critical infrastructure risks within high-performance computing environments. Security researchers detailed CVE-2026-47483, a high-severity vulnerability discovered in Nvidia’s Data Center GPU Manager (DCGM) Exporter monitoring tool. The flaw allows unauthenticated, remote attackers to overwhelm profiling endpoints with malicious requests, triggering resource exhaustion that can crash the service and degrade or halt AI workloads running on the host system.
During internet-wide scans conducted between March and May 2026, researchers identified approximately 2,100 hosts exposing the DCGM exporter directly to the public internet without authentication, inadvertently leaking live performance telemetry from more than 12,000 advanced GPUs. Nvidia has issued security updates addressing the flaw, urging system administrators to upgrade installations to version 4.8.2 or later immediately.
Shai-Hulud-Style Worm Slips Into Tensorlake’s NPM SDK
The software supply chain faced yet another malicious injection when version 0.5.144 of the tensorlake package—the official npm SDK for Tensorlake’s AI agent sandboxes—was compromised. Security researchers from Socket and Sonatype categorized the incident as a ChainDrop or Shai-Hulud-style supply chain attack, designed to execute a credential-harvesting worm automatically during the package installation process.
The embedded malware is engineered to sweep developer systems for sensitive credentials, including npm tokens, GitHub personal access tokens, AWS keys, Kubernetes configurations, HashiCorp Vault secrets, and local AI coding tool configurations. Furthermore, the payload possesses the capability to execute arbitrary code provided by the remote attacker and can automatically republish compromised versions through other developer packages under the victim’s control.
