Databricks Tested a Stronger Model Against Its Multi-Step Agent on Hybrid Queries. The Stronger Model Still Lost by 21%.
Databricks’ research shows its multi-step Supervisor Agent beats single‑turn retrieval‑augmented generation (RAG) models on hybrid queries, delivering 20%‑plus gains on the STaRK benchmark and a 21% advantage on academic tasks even when using a stronger foundation model. The study attributes the performance gap to architectural limitations rather than model quality, emphasizing the need for agents that can simultaneously query structured SQL warehouses and unstructured vector stores. By decomposing queries, self‑correcting, and using declarative source descriptions, the agent handles complex enterprise questions without custom code. The findings suggest a shift toward tool‑oriented AI agents for data‑rich businesses.
43% of AI-Generated Code Changes Need Debugging in Production, Survey Finds
A Lightrun survey of 200 senior SRE and DevOps leaders finds that 43% of AI‑generated code changes still require manual debugging in production, even after QA and staging. Engineers are spending roughly 38% of their work week—about two full days—on...
Agentic Coding at Enterprise Scale Demands Spec-Driven Development
AWS’s Kiro platform demonstrates that spec‑driven development can shrink enterprise software cycles dramatically, turning multi‑week feature builds into multi‑day sprints. By anchoring AI agents to rich, structured specifications, teams can generate code, run property‑based tests, and let agents self‑correct without...
Is Anthropic 'Nerfing' Claude? Users Increasingly Report Performance Degradation as Leaders Push Back
Developers and AI power users have flooded GitHub, X and Reddit with complaints that Anthropic’s Claude Opus 4.6 and Claude Code have become slower, more token‑heavy and less reliable. The most detailed allegation came from AMD senior director Stella Laurenzo, who analyzed...
Designing the Agentic AI Enterprise for Measurable Performance
EdgeVerve outlines a production‑grade framework for deploying semi‑autonomous AI agents across enterprise workflows. It stresses starting with business outcomes, decomposing tasks, and building a governed, observable platform that balances autonomy with risk. A finance pilot delivered over $32 million cash‑flow lift,...
Five Signs Data Drift Is Already Undermining Your Security Models
Data drift occurs when the statistical profile of inputs to a security‑focused machine‑learning model changes, eroding its detection accuracy. The article outlines five practical signs—performance drops, distribution shifts, altered prediction patterns, rising uncertainty, and broken feature relationships—that indicate drift is...
Your Developers Are Already Running AI Locally: Why On-Device Inference Is the CISO’s New Blind Spot
The rise of on‑device large language model inference is turning the CISO’s focus from cloud‑based data exfiltration to hidden risks on employee laptops. Advances in consumer‑grade accelerators, mainstream quantization, and frictionless model distribution now let engineers run 70‑billion‑parameter models locally...
AI Agent Credentials Live in the Same Box as Untrusted Code. Two New Architectures Show Where the Blast Radius Actually...
At RSAC 2026, four security leaders warned that AI agents still operate in monolithic containers where credentials sit alongside executable code, creating a massive blast radius. New architectures from Anthropic and Nvidia aim to impose zero‑trust controls: Anthropic’s Managed Agents split...
OpenAI Introduces ChatGPT Pro $100 Tier with 5X Usage Limits for Codex Compared to Plus
OpenAI unveiled a $100 ChatGPT Pro tier that delivers five‑times the Codex usage limits of the $20 Plus plan, including a temporary 2× boost through May 31 2026. The new tier expands local‑message and cloud‑task caps across GPT‑5.4, GPT‑5.4‑mini, and GPT‑5.3‑Codex models, while the...
Mythos Autonomously Exploited Vulnerabilities that Survived 27 Years of Human Review. Security Teams Need a New Detection Playbook
Anthropic’s Claude Mythos Preview autonomously uncovered a 27‑year‑old OpenBSD TCP stack bug and dozens of other zero‑day flaws across operating systems, browsers, and crypto libraries, costing roughly $20,000 per discovery campaign. The model demonstrated a 90‑fold improvement over Claude Opus...
New Framework Lets AI Agents Rewrite Their Own Skills without Retraining the Underlying Model
Researchers introduced Memento‑Skills, a framework that lets autonomous AI agents rewrite and expand their own skill libraries without retraining the underlying large language model. By treating skills as structured markdown artifacts stored in an external memory, the system can iteratively...
LLM-Referred Traffic Converts at 30-40% — and Most Enterprises Aren't Optimizing for It
The rise of AI agents is reshaping web discovery, giving birth to Answer Engine Optimization (AEO) where content must be understood and cited by large language models rather than clicked by humans. Enterprises that continue to rely solely on traditional...
Block Introduces Managerbot, a Proactive Square AI Agent and the Clearest Proof Point yet for Jack Dorsey’s AI Bet
Block unveiled Managerbot, an AI agent embedded in Square that proactively monitors small‑business metrics and suggests actions across inventory, staffing and marketing. Built on OpenAI’s GPT and Anthropic’s Sonnet models, the tool leverages Block’s proprietary “agent harness” to coordinate hundreds...
Amazon S3 Files Gives AI Agents a Native File System Workspace, Ending the Object-File Split that Breaks Multi-Agent Pipelines
Amazon announced S3 Files, a service that mounts any S3 bucket directly into an agent’s local environment using Elastic File System technology. The solution provides true file‑system semantics while keeping S3 as the system of record, eliminating the need for...
Anthropic Says Its Most Powerful AI Cyber Model Is Too Dangerous to Release Publicly — so It Built Project Glasswing
Anthropic unveiled Project Glasswing, pairing its unreleased frontier AI model Claude Mythos Preview with a coalition of twelve leading tech and finance firms to hunt and patch critical software vulnerabilities. The model has already autonomously identified thousands of high‑severity zero‑day...
How MassMutual and Mass General Brigham Turned AI Pilot Sprawl Into Production Results
MassMutual and Mass General Brigham (MGB) tackled rampant AI pilot sprawl by imposing disciplined governance, turning experimental projects into production‑grade solutions. MassMutual reported a 30% boost in developer productivity, help‑desk resolution dropping from 11 minutes to one, and customer calls...
OCSF Explained: The Shared Data Language Security Teams Have Been Missing
The Open Cybersecurity Schema Framework (OCSF) is emerging as a de‑facto standard for describing security events, findings, and context across vendors. Since its 2022 launch, the community has expanded to roughly 900 contributors after joining the Linux Foundation, and major...
Karpathy Shares 'LLM Knowledge Base' Architecture that Bypasses RAG with an Evolving Markdown Library Maintained by AI
Andrej Karpathy unveiled an "LLM Knowledge Base" that lets a large language model act as a librarian, continuously compiling, linking, and linting markdown files instead of relying on vector databases and retrieval‑augmented generation. The workflow ingests raw research assets into...
Nvidia Launches Enterprise AI Agent Platform with Adobe, Salesforce, SAP Among 17 Adopters at GTC 2026
Nvidia unveiled its open‑source Agent Toolkit at GTC 2026, a unified software stack for building autonomous enterprise AI agents. The platform, which includes Nemotron models, the AI‑Q cost‑saving blueprint, OpenShell security runtime and cuOpt optimization libraries, is already pledged by...
Meta's New Structured Prompting Technique Makes LLMs Significantly Better at Code Review — Boosting Accuracy to 93% in some Cases
Meta researchers unveiled a "semi-formal reasoning" prompting technique that structures LLM outputs as logical certificates, compelling the model to state premises, trace execution paths, and derive conclusions before answering. In benchmark tests on patch equivalence, fault localization and code Q&A,...
OpenClaw Has 500,000 Instances and No Enterprise Kill Switch
OpenClaw, an AI‑driven personal assistant, has exploded to roughly 500,000 internet‑facing instances, with more than 30,000 showing clear security gaps. A UK CEO’s unencrypted OpenClaw workspace was listed for sale on BreachForums, exposing conversations, production databases, API keys and personal...
When Product Managers Ship Code: AI Just Broke the Software Org Chart
AI agents have reduced the cost of turning intent into working software to near‑zero, allowing product managers and designers to build and ship features directly. This eliminated traditional tickets, handoffs, and lengthy sprint cycles, collapsing cycle times from weeks to...
IndexCache, a New Sparse Attention Optimizer, Delivers 1.82x Faster Inference on Long-Context AI Models
Researchers from Tsinghua University and Z.ai introduced IndexCache, a sparse‑attention optimizer that cuts up to 75% of redundant indexer computation in DeepSeek Sparse Attention (DSA) models. The technique delivers a 1.82× speedup in time‑to‑first‑token and a 1.48× boost in generation...
Intercom's New Post-Trained Fin Apex 1.0 Beats GPT-5.4 and Claude Sonnet 4.6 at Customer Service Resolutions
Intercom unveiled Fin Apex 1.0, a purpose‑built AI model that powers its Fin customer‑service agent handling over two million weekly conversations. Benchmarks show a 73.1% resolution rate, edging out OpenAI’s GPT‑5.4 and Anthropic’s Claude models by roughly two percentage points....
Three Ways AI Is Learning to Understand the Physical World
Large language models struggle with physical causality, driving a surge in "world model" research and billion‑dollar funding rounds from AMI Labs and World Labs. Researchers are exploring three architectural families—JEPA’s latent, real‑time embeddings, Gaussian‑splat 3D scene generators, and end‑to‑end generative...
Scale AI Launches Voice Showdown, the First Real-World Benchmark for Voice AI — and the Results Are Humbling for some...
Scale AI unveiled Voice Showdown, a real‑world, preference‑based benchmark for voice AI that lets users converse with frontier models for free and vote on the better response. The platform captures thousands of spontaneous interactions in over 60 languages, revealing gaps...
Why Enterprises Are Replacing Generic AI with Tools that Know Their Users
Enterprises are moving beyond generic AI models toward agents that understand individual users, leveraging large language models for deep personalization. Zoom exemplifies this shift with its AI Companion, which lets users tailor meeting summaries, generate persona‑specific follow‑up emails, and apply...
New MiniMax M2.7 Proprietary AI Model Is 'Self-Evolving' And Can Perform 30-50% of Reinforcement Learning Research Workflow
MiniMax unveiled its proprietary M2.7 large language model, a reasoning‑only LLM that can autonomously manage 30‑50% of its own reinforcement‑learning development loop. The model earned a 66.6% medal rate on the MLE Bench Lite, tying Google Gemini 3.1 and approaching Anthropic...
Enterprise AI Agents Keep Operating From Different Versions of Reality — Microsoft Says Fabric IQ Is the Fix
Microsoft announced a major upgrade to its Fabric IQ semantic layer, making the business ontology accessible via the Machine‑Controlled Protocol (MCP) to any AI agent, regardless of vendor. The update also adds enterprise planning capabilities and introduces a Database Hub...
Mistral AI Launches Forge to Help Companies Build Proprietary AI Models, Challenging Cloud Giants
Mistral AI unveiled Forge, an enterprise‑grade model training platform that lets organizations build, fine‑tune, and continuously improve AI models using their own proprietary data. The service covers the full training lifecycle—from pre‑training on large internal datasets to reinforcement‑learning alignment—running on...
Nvidia's Agentic AI Stack Is the First Major Platform to Ship with Security at Launch, but Governance Gaps Remain
Nvidia unveiled its agentic AI stack at GTC, marking the first major AI platform to ship with security baked in rather than added later. Five security vendors—CrowdStrike, Palo Alto Networks, JFrog, Cisco, and World Wide Technology—each cover a distinct layer...
The Accessibility Gap: Why Good Intentions Aren’t Enough for Digital Compliance
AudioEye’s 2026 Accessibility Advantage Report reveals a stark gap between businesses’ awareness of digital accessibility and their ability to execute it. While 59% of leaders acknowledge legal risk and more than half have faced lawsuits, the average web page still...
Rethinking AEO when Software Agents Navigate the Web on Behalf of Users
The rise of AI‑powered agents that browse the web on users' behalf is eroding the long‑standing assumption that every click, scroll or purchase funnel step reflects a conscious human decision. While the raw data—page views, button clicks, time on page—remains...
NanoClaw and Docker Partner to Make Sandboxes the Safest Way for Enterprises to Deploy AI Agents
NanoClaw has partnered with Docker to run its open‑source AI agent platform inside Docker Sandboxes, providing enterprise‑grade isolation for autonomous agents. The integration leverages MicroVM‑based sandboxes, allowing agents to install packages, modify files, and access external systems without exposing the...
Y Combinator-Backed Random Labs Launches Slate V1, Claiming the First 'Swarm-Native' Coding Agent
Random Labs, a YC‑backed startup, launched Slate V1, the first “swarm‑native” autonomous coding agent. Slate uses a dynamic pruning algorithm and a Thread Weaving architecture to orchestrate parallel worker threads, preserving context across large codebases. The system separates strategic orchestration from...
Agents Need Vector Search More than RAG Ever Did
Qdrant, a Berlin‑based open‑source vector search firm, announced a $50 million Series B round and launched platform version 1.17. The update introduces relevance‑feedback queries, delayed fan‑out, and cluster‑wide telemetry to support the high‑throughput demands of autonomous AI agents. It highlights that agents...
Manufact Raises $6.3M as MCP Becomes the ‘USB-C for AI’ Powering ChatGPT and Claude Apps
Manufact, a YC‑backed startup, announced a $6.3 million seed round led by Peak XV to build infrastructure for the Model Context Protocol (MCP), the emerging “USB‑C” standard for AI agents. The company’s open‑source mcp‑use SDK has already logged five million downloads and...
RSAC's Innovation Sandbox Is Where Cybersecurity's Next Giants Are Born
The RSAC Innovation Sandbox celebrates its 20th year, showcasing ten cybersecurity startups tackling AI governance, identity, and supply‑chain risks. Over the past two decades the contest has spurred more than $50.1 billion in investments and over 100 acquisitions among its alumni....
Google Finds that AI Agents Learn to Cooperate when Trained Against Unpredictable Opponents
Google’s Paradigms of Intelligence team demonstrated that training large‑language‑model agents with decentralized reinforcement learning against a mixed pool of static and learning opponents produces cooperative behavior without hard‑coded rules. The agents use in‑context learning to infer co‑player strategies in real...
How to Make Your E-Commerce Product Visible to AI Agents? Use This New System Trusted by L’Oréal, Unilever, Mars &...
Azoma has launched the Agentic Merchant Protocol (AMP), a machine‑native framework that lets brands push product data directly to AI shopping agents. The system, already adopted by L’Oréal, Unilever, Mars and Beiersdorf, centralizes catalogs, enforces brand guidelines and promises measurable...
Enterprise Agentic AI Requires a Process Layer Most Companies Haven’t Built
Enterprises are racing to adopt agentic AI, with 85% targeting deployment within three years, yet 76% admit their current operations lack the necessary process foundation. Celonis’ 2026 Process Optimization Report reveals that only 19% of firms have implemented multi‑agent systems,...
LangChain's CEO Argues that Better Models Alone Won't Get Your AI Agent to Production
LangChain CEO Harrison Chase argues that superior large language models alone won’t drive AI agents into production; the missing piece is sophisticated "harness engineering" that lets models run loops, call tools, and manage context. He highlights the evolution from static...
Karpathy’s March of Nines Shows Why 90% AI Reliability Isn’t Even Close to Enough
Andrej Karpathy’s “March of Nines” highlights that achieving 90% AI reliability is only the first step; each additional nine of uptime demands comparable engineering effort. In multi‑step agentic workflows, the probability of success compounds exponentially, turning a seemingly robust demo...
Anthropic Launches Claude Marketplace, Giving Enterprises Access to Claude-Powered Tools From Replit, GitLab, Harvey and More
Anthropic unveiled Claude Marketplace, a limited‑preview platform that lets enterprises apply existing Claude spend toward third‑party tools built on its models. The offering aggregates partners such as GitLab, Replit, Snowflake, Harvey and Rogo, and consolidates invoicing to streamline procurement. By...
New KV Cache Compaction Technique Cuts LLM Memory 50x without Accuracy Loss
MIT researchers introduced Attention Matching, a fast KV‑cache compression technique that can shrink large language model memory footprints by up to 50 times without measurable accuracy loss. The method preserves attention output and attention mass by fitting compressed keys and values...
Google PM Open-Sources Always On Memory Agent, Ditching Vector Databases for LLM-Driven Persistent Memory
Google has open‑sourced an Always On Memory Agent that eliminates traditional vector‑database retrieval in favor of a large‑language‑model‑driven memory layer. Built on the Agent Development Kit and powered by the low‑cost Gemini 3.1 Flash‑Lite model, the agent continuously ingests, consolidates, and serves...
Databricks Built a RAG Agent It Says Can Handle Every Kind of Enterprise Search
Databricks unveiled KARL, a reinforcement‑learning‑driven RAG agent that tackles six distinct enterprise search behaviors in a single model. The system matches Claude Opus 4.6 on the proprietary KARLBench benchmark while delivering 33% lower cost per query and 47% reduced latency. KARL...
Pentagon Vendor Cutoff Exposes the AI Dependency Map Most Enterprises Never Built
The Pentagon’s six‑month ban on Anthropic’s Claude has exposed a blind spot in enterprise AI risk management: most firms cannot map the full chain of AI model dependencies. A Panorays survey shows only 15% of CISOs have complete visibility, while...
Did Alibaba Just Kneecap Its Powerful Qwen AI Team? Key Figures Depart in Wake of Latest Open Source Release
Alibaba released the Qwen3.5 small model series, praised for its high intelligence density and ability to run on consumer devices. Within 24 hours, technical architect Junyang "Justin" Lin and two teammates announced their departures, leaving the future of the Qwen...
EY Hit 4x Coding Productivity by Connecting AI Agents to Engineering Standards
EY’s product development team boosted coding productivity four‑ to five‑fold by wiring AI coding agents into its engineering standards, code repositories, and compliance frameworks. The initiative required an 18‑ to 24‑month effort to embed cultural acceptance and technical integrations, moving...