Technology worth knowing today.
Artificial Intelligence
Google unveils Gemini 3.8 Live with Live Avatar feature
What happened. Google unveiled Gemini 3.8 Live, the latest iteration of its Gemini language model family, highlighting a new Live Avatar feature that enables real‑time visual representation of the model’s responses. The announcement positions the release as an update to the existing Gemini lineup, focusing on multimodal interaction without detailing pricing, availability, or benchmark results. The Live Avatar is presented as a way to animate model output synchronously with generated text or speech.
Why it matters. The Live Avatar addition gives developers a built‑in method to pair model output with a synchronized visual character, reducing the need for separate animation pipelines when building interactive agents, virtual assistants, or training simulations. By integrating the avatar directly into the model’s output stream, teams can prototype multimodal experiences faster and maintain tighter coupling between spoken or textual responses and visual feedback, which may simplify workflows for applications that rely on real‑time user engagement.
Who should care. Developers · Cloud architects · Platform engineers · Technology leaders · People learning AI · Students and career changers
Source: Google DeepMind
Scribd scales trust‑and‑safety classification of 400M+ documents using Gemini Enterprise batch inference
What happened. Scribd used Gemini Enterprise batch inference to classify over 400 million user-uploaded documents spanning more than 12 billion pages across its Scribd and Slideshare platforms. By feeding native PDFs directly into the model, the company avoided OCR or rendering pipelines and completed the trust‑and‑safety backfill in a few months, leveraging Google Cloud’s scalable batch throughput. Gemini Enterprise’s batch pricing, offered at a 50 % discount to interactive rates, made large‑scale LLM classification economically viable for the entire corpus.
Why it matters. This demonstrates that a production-grade LLM can be run at corpus scale without heavy preprocessing, showing how native PDF handling and batch-mode pricing lower the operational barrier for large-scale content moderation. For teams managing large document repositories, the approach offers a template for integrating AI safety checks while keeping compute costs predictable, and it highlights the role of cloud-batch services in meeting tight timelines for AI-driven data processing.
Who should care. Developers · Cloud architects · Security teams · Data engineers · Technology leaders
Source: Google Cloud
Agent Harnesses Explained: How LLMs Gain Live‑Data Autonomy
What happened. In a recent Agent Factory episode, Ryan Lopopolo of Google Cloud explained that an AI agent consists of a large language model plus an agent harness, which supplies the model with live data and workspace interactions. He described how moving interventions earlier in the development process—shifting left—enables agents to operate more autonomously, using tools like Google Antigravity with Gemini 3.8 Flash to answer context‑dependent questions.
Why it matters. By wrapping LLMs in a harness that can query live services and feed results back into the prompt, agents gain the ability to act on current information without manual intervention. This approach reduces reliance on static model knowledge, lets developers delegate routine checks to software, and moves complex decision‑making earlier in the workflow, which can improve efficiency and reduce errors in AI‑driven applications.
Who should care. Developers · Cloud architects · Platform engineers · Technology leaders · People learning AI
Source: Google Cloud
Proaction Reports 60% Sales Boost and 75‑Hour Savings Using AI Coding Tools
What happened. Proaction says that using Codex alongside GPT-Live-1 and GPT-6 Astra helped the company increase sales by 60% and cut more than 75 hours of work, according to the companys own report. The AI tools were applied to the development, operation, and sales processes for its fleet-management platform. The team said the automation reduced manual coding tasks and accelerated feature releases, leading to shorter time-to-market.
Why it matters. If accurate, the result would suggest that generative-AI coding assistants can translate into concrete business outcomes when integrated into product engineering workflows. However, these figures come solely from Proactions own self-reported case study and have not been independently verified, so technology leaders should treat them as a vendor-supplied data point rather than an established benchmark.
Who should care. Developers · Cloud architects · Platform engineers · Technology leaders
Source: OpenAI
Google Cloud releases Memorystore for Valkey 9.1 with up to 3x QPS and microsecond latency for AI workloads
What happened. Google Cloud announced the general availability of Memorystore for Valkey 9.1, a managed in-memory data store based on the open‑source Valkey fork of Redis. The release delivers up to three times the queries per second of the existing Memorystore for Redis Cluster while maintaining microsecond latency, thanks to a redesign of thread communication that offloads I/O work from the main execution loop. Valkey, governed by the Linux Foundation, is positioned as a performant option for AI‑driven microservices that need to serve millions of concurrent users.
Why it matters. The higher throughput and low latency enable AI applications that rely on fast caching—such as retrieval‑augmented generation, real‑time recommendation engines, or agent‑based workflows—to handle more requests without adding infrastructure. By providing a managed, open‑source alternative to Redis, Google Cloud reduces vendor lock‑in risk while giving teams a proven path to scale AI workloads that demand millions of concurrent interactions. This addresses a common bottleneck in AI‑centric architectures where data access speed directly impacts model responsiveness and cost efficiency.
Who should care. Developers · Cloud architects · Platform engineers · Technology leaders · Data engineers
Source: Google Cloud
Google Cloud releases Storage Intelligence advisor and batch operations to streamline storage governance
What happened. Google Cloud announced the general availability of Storage Intelligence advisor and expanded storage batch operations for all Storage Intelligence customers. The advisor analyzes changes across a storage estate and recommends actions, while batch operations can apply those recommendations to millions of objects. The features aim to reduce the manual effort of exporting inventory, joining logs, and building dashboards, addressing the growing complexity of AI‑driven data workloads that outpace traditional governance.
Why it matters. By surfacing what changed in storage and offering automated remediation at scale, the advisor and batch tools let organizations spot unusual access patterns or data growth before they appear on bills, cutting the need for custom inventory pipelines. This reduces operational overhead and helps maintain cost predictability as AI training and inference generate rapidly shifting data volumes.
Who should care. Cloud architects · Platform engineers · Data engineers · Administrators · Technology leaders
Source: Google Cloud
Cloud Computing
New “no‑compromises” database architecture targets agentic workloads
What happened. The source discusses how existing database architectures—such as Exadata, Azure SQL Hyperscale, Aurora, and object‑storage‑backed designs—each face trade‑offs among scale, latency, and isolation. It argues that in the agentic era, where workloads are generated dynamically and cannot be pre‑vetted, those compromises are unacceptable because isolation from mission‑critical systems becomes a business‑continuity requirement. The article then introduces a new database architecture described as “no‑compromises” for the agentic era, aiming to overcome those limitations.
Why it matters. The need for isolation in agentic workloads means that database platforms must provide strong workload separation without sacrificing performance or scalability; otherwise mission‑critical data could be affected by unpredictable agentic traffic. A design that eliminates the scale‑latency‑isolation trade‑off would let developers run AI‑driven agents alongside transactional workloads on the same platform, reducing operational complexity and the need for separate data stores. This matters for cloud architects and platform engineers who must support evolving AI workloads while maintaining SLAs.
Who should care. Developers · Cloud architects · Platform engineers · Data engineers · Technology leaders · People learning AI
Source: Google Cloud
DevOps
First‑time KubeCon speaker shows how to run an AI cluster on a DGX Spark with Kubernetes and DRA
What happened. At KubeCon plus CloudNativeCon India 2026 in Mumbai, a first-time conference speaker presented alongside their father on turning an NVIDIA DGX Spark into a self-hosted AI cluster. The talk walked through Kubernetes setup, exposing GPUs to workloads, and using Dynamic Resource Allocation to schedule jobs, all on hardware the presenters owned and controlled themselves rather than renting from a cloud provider. A recording of the session is posted on YouTube.
Why it matters. Running the full AI stack on owned hardware cuts a teams dependence on proprietary cloud services and puts hardware choices, security posture and cost directly in their own hands. The talk sketches a repeatable blueprint — Kubernetes plus GPU access plus Dynamic Resource Allocation on a single DGX Spark node — that platform teams and developers could adapt to build their own on-premises AI clusters.
Who should care. Developers · Cloud architects · Platform engineers · Administrators · Technology leaders
Source: CNCF
Cybersecurity
Norwegian Cruise Line Door Controllers Flaw Allows RFID Cloning
What happened. Norwegian Cruise Lines door access controllers on its ships contain an improper authentication flaw (CVE-2026-75907) that relies only on the static 7-byte UID of RFID keycards. The UID is not secret and the readers do not verify any cryptographic data, so an attacker who can briefly approach a valid card can copy its UID onto a writable card and gain entry to secured areas. The vulnerability allows cloning of credentials without needing to alter the original card.
Why it matters. The flaw shows how reliance on a simple identifier for access control can be defeated by low‑cost RFID cloning tools, highlighting a common pitfall in physical security systems that use NFC or RFID without cryptographic verification. For developers and administrators, it underscores the need to authenticate credentials with challenge‑response mechanisms rather than static IDs, especially in environments like cruise ships where physical proximity to legitimate tokens is easy for attackers.
Who should care. Developers · Security teams · Administrators · Technology leaders · Platform engineers
Source: CERT/CC
Check Point Products Affected by Path‑Traversal Flaw Added to CISA KEV Catalog
What happened. Check Point disclosed a path-traversal vulnerability, tracked as CVE-2026-93616, affecting its Security Management Server, Multi-Domain Security Management Server, Log Server, Multi-Domain Log Server and SmartEvent products. CISA has added the flaw to its Known Exploited Vulnerabilities catalog. An attacker needs no credentials to trigger the bug, which stems from how these products resolve file paths, and successful exploitation lets them upload and execute arbitrary scripts on the affected system.
Why it matters. Since no authentication is needed, anyone who can reach a vulnerable Check Point management interface over the network could upload and run code of their choosing on the server. That risk cuts through normal authentication controls and could let an attacker tamper with logs, rewrite security policy, or establish a longer-term foothold in the environment.
Who should care. Security teams · Administrators · Developers · Platform engineers
Source: CISA (catalog reference)
Cisco Talos releases CAIRN toolkit to hunt AI‑integrated malware via metadata artifacts
What happened. Cisco Talos has unveiled CAIRN (Cognitive Artifact Intelligence Research Network), a research toolkit that Talos says can help detect AI-integrated malware by examining metadata such as prompt templates, provider endpoints, API keys and jailbreak terms rather than the binary itself. The approach treats these strings as cognitive artifacts left by attackers, and according to Talos enables a metadata-first hunting methodology intended to be fast and scalable. Initial findings will be shared starting with a sample called CLOSEDQUORUM.
Why it matters. According to Cisco Talos, focusing on artifacts left during AI integration could offer defenders a way to spot malware that leverages AI systems without needing to reverse-engineer the payload. If borne out, this approach may help security teams prioritize hunting efforts and reduce reliance on binary analysis alone — though the claim comes from Talos own announcement and has not been independently verified.
Who should care. Security teams · Developers · Platform engineers · Administrators · Technology leaders · People learning AI
Source: Cisco Talos
Kiteworks Warns Customers to Halt Use After Federal Threat Alert
What happened. Kiteworks told its customers to stop using its platform after receiving credible threat intelligence from U.S. federal intelligence agencies. The intelligence indicated that a threat actor might try to target some Kiteworks systems belonging to customers. Frank Balonis, the company’s CISO, shared the warning with Recorded Future News, saying the alert came from federal authorities and suggested a possible targeting attempt. The company did not disclose which specific systems or customers could be affected.
Why it matters. The alert shows that a widely used collaboration platform is flagged by federal intelligence as a possible target, which could put customer data and workflows at risk if the threat materializes. It underscores the value of sharing threat intelligence between government and private sector, and reminds organizations that even trusted services can be caught in broader cyber‑espionage campaigns, prompting a review of access controls and monitoring.
Who should care. Security teams · Administrators · Platform engineers · Technology leaders · Cloud architects
Source: The Record
Labcorp fined $2.3 million, vows to tighten vendor data security
What happened. Labcorp agreed to pay a $2.3 million fine after regulators cited cybersecurity shortcomings in its handling of vendor data. The company said it will overhaul its data‑security program by creating an incident‑response plan focused on vendor failures, limiting the volume of patient information shared with third‑party providers, and establishing a dedicated risk‑management team to monitor vendors’ compliance with security standards. Labcorp described the changes as steps to strengthen oversight and reduce future exposure to supply‑chain threats.
Why it matters. The settlement underscores how third‑party relationships can trigger regulatory penalties when vendor security lapses are not adequately managed. By requiring Labcorp to build an incident‑response plan for vendor issues, tighten data‑sharing limits, and staff a risk‑management team to oversee compliance, the case shows a growing expectation that healthcare organizations treat their supply chain as part of their own security perimeter. For security practitioners, it signals that vendor oversight, data‑minimisation practices, and continuous compliance monitoring are becoming baseline expectations rather than optional extras.
Who should care. Security teams · Administrators · Technology leaders · Developers
Source: The Record
Bitget CEO Claims North Korea Stole $387 Million, Calls on User Protection Fund to Cover Losses
What happened. The chief executive of cryptocurrency exchange Bitget said that actors linked to North Korea stole $387 million from the platform. He stated that the exchange maintains a User Protection Fund exceeding $464 million, which will be deployed to reimburse affected users. The accusation was made in a public statement, with the CEO asserting that the fund’s size is sufficient to cover the alleged losses. No independent verification of the theft or the fund’s exact balance was provided in the statement.
Why it matters. The claim underscores how even large cryptocurrency exchanges can be targeted by sophisticated, state‑linked actors, reminding security teams that traditional defenses may not stop nation‑level threats. It also shows why exchanges maintain user protection funds—to absorb losses and preserve trust when breaches occur. The incident highlights the need for continuous monitoring, incident response planning, and transparent communication about financial safeguards.
Who should care. Security teams · Administrators · Technology leaders · Developers · Platform engineers
Source: The Record