Foundation Models Push Into Labs as AI Infra and Security Heat Up

AI & Machine Learning

Anthropic will let enterprise customers exercise greater control over how long the company retains conversation and telemetry data, shifting from prior defaults to options designed for corporate privacy and compliance needs; the change is positioned as a commercial product and data‑governance update that could affect procurement and contractual terms for large buyers. This move underscores growing enterprise demand for configurable retention policies as AI systems move into regulated environments, and it may reduce one barrier for organizations hesitant to deploy conversational AI in sensitive workflows. The update could also prompt competitors to offer similar controls or clearer SLAs, changing how enterprises evaluate vendors on compliance risk. Source: Reuters Verified: True

Nvidia unveiled SONIC, a foundation‑model approach that trains a single unified model to teach humanoid robots locomotion and manipulation skills, eliminating the need for separate task‑by‑task controllers and simplifying skill transfer. The system demonstrates the ongoing convergence between large model techniques and embodied robotics, suggesting foundation models can scale across sensory and motor domains rather than remaining purely perceptual or language-focused. If SONIC’s methods generalize outside simulation, it may accelerate research on sim‑to‑real transfer and reduce engineering overhead for multi‑task robots. Source: AiBusiness Verified: True

Noitom Robotics published HiPHI, a large, high‑precision human motion dataset containing hundreds of hours of motion capture from many performers interacting with objects, aimed at accelerating reproducible work in human motion understanding and embodied AI. By providing labeled, high‑fidelity motion sequences, HiPHI targets researchers in animation, robotics, and biomechanics who need consistent training data for learning robust controllers and perception modules. Large public motion datasets can lower the barrier to entry for labs and startups building human‑centric models and improve benchmarking for control and imitation learning. Source: Business Insider Markets Verified: True

GenBio launched a “Virtual Cell” foundation model intended to simulate cellular behavior and predict perturbation effects, pitching the system as a tool to accelerate biological discovery and early drug R&D. The model applies large‑scale modeling approaches to molecular and pathway dynamics, aiming to let researchers explore hypotheses in silico before committing lab resources, which could reduce iteration time in preclinical research. While promising, the approach will require rigorous benchmarking and experimental validation before it can be relied on for high‑stakes decisions in drug development. Source: Lifespan.io Verified: True

Consumer Hardware

No major stories this sector today.

Cybersecurity

Security researchers at Wiz disclosed a GitHub Actions workflow‑injection vulnerability in a Snowflake integration that can allow attackers to craft repository issues which then trigger command injection paths in CI workflows, potentially enabling code execution or secrets exposure. The issue highlights a supply‑chain and CI/CD vector risk for organizations using Snowflake’s public GitHub integrations, as attackers can weaponize common developer workflows to escalate access. Wiz and downstream vendors have published mitigation guidance, and organizations should review affected workflows, apply recommended patches, and restrict repository issue‑triggered actions until fixes are verified. Source: The Hacker News Verified: True

Researchers at Malwarebytes disclosed “ShieldBreak,” a technique that bypasses Microsoft’s earlier patch for a Defender elevation‑of‑privilege vulnerability and can allow attackers to regain SYSTEM privileges under some configurations. The writeup details exploitation paths and recommends additional mitigations beyond the initial patch, signaling sustained exploitation pressure around Windows endpoint protections and the need for layered defenses. Organizations should follow vendor guidance, apply the extra mitigations, and monitor for indicators of compromise while emergency response teams assess impacted environments. Source: Malwarebytes Verified: True

Enterprise Infrastructure

Velaura AI raised $110 million in a Series A round that values the chip‑design startup at over $1 billion, underscoring investor appetite for custom silicon and IP optimized for AI workloads. The company focuses on accelerators tailored to training and inference cost efficiencies, reflecting a market that still sees differentiated silicon design as a pathway to lower operational AI spend. The funding signals continued capital flow into startups building alternatives to hyperscaler designs and may accelerate partnerships with foundries or systems integrators. Source: Reuters Verified: True

Velatir, an Odense‑based startup building combined software and hardware stacks to run models near data and edge locations, secured €5 million to accelerate European expansion and product development for AI runtime and infrastructure tooling. The round highlights demand for regionalized AI compute that reduces latency and addresses data‑sovereignty requirements, particularly for European customers wary of centralized cloud models. Expanding locally distributed runtimes could ease compliance hurdles and open new markets for low‑latency, privacy‑sensitive applications. Source: EU‑Startups Verified: True

Twin A1, a digital‑twin startup targeting professional services, raised $20 million in seed funding to commercialize products that capture domain experts’ knowledge and decision patterns for B2B workflows in legal, consulting, and regulated industries. The platform aims to embed expertise and workflow automation into enterprise systems, enabling knowledge retention and scaling of advisory services through digital twins of processes and decisions. If successful, these products could change how firms package expertise, automate routine tasks, and guard institutional knowledge. Source: Law.com Verified: True

Pixalate launched a pre‑bid IPv6 Data‑Center Blocklist intended to extend invalid‑traffic and ad‑fraud protections across IPv6 address ranges for roughly 3,000 cloud providers, allowing programmatic buyers to block known data‑center and bot traffic before bidding. The product targets programmatic ad platforms and enterprises that need to scale fraud mitigation into IPv6 environments as cloud providers expand their address space. Pre‑bid filtering at scale can reduce wasted spend and privacy risk, though it depends on accurate attribution and ongoing blocklist maintenance. Source: GlobeNewswire Verified: True

Policy & Regulation

No major stories this sector today.