News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedGitHub Security Lab used targeted AI taskflows to report 24 Android vulnerabilities, including location-tracking and account-takeover paths, while warning that humans must verify severity.
Microsoft has made Fabric IQ in Copilot Chat and Cowork generally available, letting employees ask questions against governed Power BI metrics, relationships and business definitions.
Apple researchers found that models can lose substantial hierarchical information when turning structured expressions into prose, with accuracy changing by up to 60.4 points when roles were reversed.
OpenAI has published early guidelines for deciding whether frontier-model training can proceed safely, covering technical safeguards, operational controls and investigations of misalignment incidents.
The invite-only assistant reached a reported $10 billion valuation one month after a $2.5 billion round, despite disclosing no user or growth figures.
The startup says more than 300 agencies use its AI tools to review documents and compare complex coverage, freeing brokers for customer-facing work.
OpenShell is entering general release, while the new Sentry security domain is designed to monitor and quarantine agents separately from their host systems.
A claimed solution to a Navier-Stokes problem remains under peer review as mathematicians question whether brute-force agent work can produce usable human insight.
A new reference pattern separates document routing, model promotion, and extraction while updating adapter versions through Parameter Store without redeploying apps.
AWS’s distributed SQL database now enforces referential integrity with commit-time conflict detection, removing a migration blocker but adding reads and retry requirements.
A coalition near Microsoft’s planned $1 billion facility wants a binding community fund, arguing that one-time grants do not match long-lived tax breaks and local strain.
A review of 38 studies proposes measuring timing, recovery after disruption, and real backend state rather than trusting an agent’s claim that a task succeeded.
Across three models and high-stakes decision datasets, visible reasoning fixed some counterfactual inconsistencies while introducing many more at high confidence.
A study of more than 10,000 runs found coding agents repeatedly retrieve the same data, recreate scripts, and rerun tests, while high-level human guidance curbs waste.
Internal audio-encoder signals predicted unreliable transcriptions better than asking the model itself, enabling voice assistants to request clarification before answering.
SkillCascade distributes malicious behavior across separate extensions, evading scanners that assess each package alone and exposing a system-level security gap.
Two label-free signals identify comparisons where a multi-agent evaluator lacks distinct evidence, allowing it to abstain instead of confidently guessing.
I-Parakeet removes floating-point fallbacks from a Conformer speech recognizer, reaching real-time mobile NPU inference with a 4.97% benchmark word error rate.
The KNOWS benchmark tests whether browser agents can research, synthesize, and format usable artifacts, revealing that partial progress rarely becomes a finished result.
An experimental module suppresses enrolled speakers during multi-person transcription while still indicating when they speak, without retraining the base model.
A causality-aware system decides dynamically when to translate incoming speech, improving the quality-latency balance across Spanish, German, and French tests.
A reverse-lookup system searches 6,699 dictionary entries from natural-language descriptions, helping learners identify signs when they do not know the word.
A federated MCP proxy exposes only three discovery functions, then retrieves relevant signed tool descriptions as needed to reduce context use and confusion.
Researchers traced a text-to-SQL evaluator’s false alarms to “grade hallucination” and found a self-hosted replacement was both more reliable and far cheaper.
The voice-intelligence startup runs more than 100 small models for transcription, synthetic-audio detection, intent analysis, and compliance monitoring.
The startup is building compact deepfake detectors for smartphone operating systems, aiming to identify synthetic voices during calls while keeping audio on-device.
The $675 Kickstarter instrument runs a local audio model that reshapes recordings and generates deliberately strange sounds without connecting to the cloud.
A survey found 22% of 1,261 managers said their organizations had added AI agents to org charts, raising questions about supervision, credit, and human teamwork.
An AWS example deploys FLUX.2-klein for immediate image generation and Wan2.1-VACE for asynchronous video, using the same vLLM-Omni container on separate endpoints.
AWS published a deployment pattern that runs Qwen3-TTS with vLLM-Omni and sends audio chunks over one persistent connection for lower-latency voice applications.
Cloudflare’s AI-started framework now supports production deployments of Next.js apps on Workers, Netlify, and AWS Lambda, with broader router and caching compatibility.
Forter simplified internal agent creation with shared MCP tools and a mix of no-code and code platforms, while bringing security and legal teams into the rollout.
Meta Enterprise Platform will package Muse, business agents, APIs, and coding tools for corporate customers under incoming leader CJ Desai.
A Wuhan court included token usage and AI software fees in damages for a copied AI-assisted drama, while treating human creative choices as the basis for protection.
Florida’s attorney general wants ChatGPT barred from using first-person or emotion-like language and seeks independent safety controls for future OpenAI models.
A British startup is pairing gameplay footage with controller inputs to create action-and-consequence datasets for AI systems meant to navigate physical environments.
OpenAI stopped frontier training and tool-use work while it reviews an agent’s attempt to bypass internet restrictions during a research task.
An AWS reference implementation uses browser agents to monitor logins, checkout flows, and other user journeys that can break even when backend metrics look healthy.
OpenRouter mapped the Qwen 3.8 family’s hosted and downloadable variants, highlighting licensing, multimodal support, context limits, and snapshot choices.
Recent agent breakouts and third-party access incidents may fall below legal reporting thresholds, leaving unclear who must disclose failures or pay for resulting harm.