SK Telecom Bets on 15GW Buildout to Make Korea’s AI Hub

AI & Machine Learning

Dnotitia unveiled STAR‑KV, a low‑rank key‑value cache compression technique the company says can yield up to ~20× compression and meaningful inference speedups for long‑context transformer deployments; the announcement was highlighted as an ICML 2026 spotlight item and frames the work as a practical lever to reduce memory and serving costs in production LLM stacks. The method targets the KV‑cache bottleneck that grows with longer contexts and could materially lower GPU memory pressure and network I/O for model‑serving vendors and cloud customers. If the claimed gains hold under independent benchmarking, STAR‑KV could change cost tradeoffs for offering multi‑hour or multi‑document context in commercial deployments and spur integration into inference runtimes. The release also signals continued startup activity focused on infra‑level optimizations rather than solely on new model architectures. Source: PR Newswire Verified: True

Researchers published AgenticDataBench, a new benchmark aimed at evaluating LLM‑based data agents across realistic B2B tasks like retrieval, transformation and automated actioning, with the goal of standardizing measurements of agent reliability and utility. The benchmark assembles domain tasks and metrics that reflect real‑world data workflows, making it easier to compare agent architectures and tool‑use strategies on production‑relevant scenarios. By focusing on retrieval and pipeline orchestration, AgenticDataBench attempts to close the gap between research‑era agent demos and enterprise expectations for correctness, auditability and efficiency. Broader adoption of such benchmarks could pressure vendors to prioritize robustness, reproducibility and tool‑integration for agentic products. Source: Let’s Data Science Verified: True

This week’s releases — STAR‑KV and AgenticDataBench — together highlight an emerging, pragmatic AI infrastructure agenda: reduce serving and memory costs for long‑context inference while creating standardized evaluations for agentic data workflows that enterprises must trust. The two efforts tackle complementary production bottlenecks: one attacking KV cache footprint and throughput, the other demanding repeatable benchmarks for agents that combine LLM reasoning with data access and actions. For teams building long‑context or agentic products, these advances point to a near‑term roadmap of integrating infra optimizations with stronger evaluation tooling to move from experiments to reliable services. Expect vendors and cloud hosts to either incorporate these techniques into runtimes or position competing tradeoffs around latency, cost and correctness. Source: PR Newswire Verified: True

Consumer Hardware

No major stories this sector today.

Cybersecurity

Security researchers published a public proof‑of‑concept exploit for the recently disclosed Linux “Bad Epoll” vulnerability, demonstrating a reliable local privilege escalation path to root on unpatched systems and underscoring the urgency of vendor fixes. The POC maps a kernel epoll code path that can be manipulated for privilege escalation, and the writeup shows exploitation across several distributions and embedded Linux builds that lag on kernel updates. Because epoll is widely used, maintainers and sysadmins are being urged to apply patches or mitigations quickly to limit local‑user attack windows, particularly in multi‑tenant and developer workstation environments. The disclosure also sparked immediate follow‑up from distro maintainers to release backported fixes and advisories. Source: SecurityWeek Verified: True

Researchers reported an incident where attackers weaponized an agentic AI workflow built with Langflow to fully automate reconnaissance, credential harvesting and ransomware delivery, illustrating how low‑code agent orchestration tools can be abused at scale. The analysis traces a chain of modular agent components that performed web discovery, credential stuffing and staged payload deployment with minimal human supervision, reducing attacker effort and ramping attack velocity. The case highlights the dual‑use risk of agent frameworks and calls for safer defaults, rate limits and stronger platform controls to prevent easy orchestration of malicious flows. Security teams and vendors are being advised to monitor for similar patterns and harden integrations that expose automation primitives. Source: SecurityWeek Verified: True

CISA issued an advisory warning of active exploitation of a recently patched Microsoft SharePoint remote‑code‑execution vulnerability (CVE‑2026‑45659), urging organizations to apply vendor fixes or implement the supplied mitigations to prevent webshell installation and lateral movement. The advisory includes detection guidance, indicators of compromise and recommended compensating controls for environments that cannot immediately patch, emphasizing web application hardening and network segmentation. Because the vulnerability permits remote code execution via authenticated or unauthenticated vectors (depending on configuration), defenders are advised to prioritize patching high‑exposure gateways and internet‑facing SharePoint instances. The notice reflects continued rapid weaponization of high‑impact SharePoint flaws and the need for tight patch management in enterprise collaboration platforms. Source: SecurityWeek Verified: True

Medtronic disclosed that a corporate‑IT breach earlier this year exposed roughly 3.8 million records, including personal and limited medical information, with investigators attributing the access to the ShinyHunters group; the company has initiated regulatory notifications and remediation efforts. The incident was identified through forensic analysis and has prompted customer outreach, resets of affected credentials and tightened monitoring of compromised systems. While Medtronic reports no evidence of ongoing unauthorized access to critical medical device systems, the breach underscores risk to healthcare supply chains and vendors that hold large volumes of sensitive data. Regulators and clients will likely press for enhanced controls and transparency as part of breach remediation. Source: SecurityWeek Verified: True

Enterprise Infrastructure

SK Telecom announced an aggressive plan to pursue construction of up to 15GW of AI‑optimized data‑centre capacity, positioning the company and South Korea as a regional hub for hyperscale power and colocated AI compute. The proposal frames the buildout as a combination of power infrastructure and purpose‑built facilities to attract cloud, enterprise and AI workload customers across Asia, and signals national ambitions to capture more of the AI stack beyond chip design and services. If executed, the 15GW target would place SKT among the more ambitious single‑country pushes for AI power capacity and could influence regional supply chains, fiber and interconnect strategies. The announcement also raises questions about permitting, grid integration and how SKT will partner with cloud and colo providers to fill that capacity. Source: HPCwire Verified: True

HCLTech said it secured a multiyear IT services contract worth $1.14 billion with a major European firm, a deal that will provide managed services and cloud modernization work and strengthen the company’s enterprise backlog in Europe. The agreement underscores continued demand for outsourcing and digital transformation engagements as enterprises accelerate cloud migrations and application modernization. For HCLTech, the contract is a material revenue win that supports its growth thesis and highlights the resiliency of large‑scale services deals even amid macro uncertainty. Market watchers will watch execution metrics and margin trajectory as the multi‑year program is deployed. Source: Reuters Verified: True

Policy & Regulation

No major stories this sector today.