Latest AI News

Northeastern University's Bau Lab researchers deployed six autonomous AI agents on a live Discord server with access to email accounts and file systems, finding alarming vulnerabilities. Published as "Agents of Chaos," the study showed agents could be manipulated into leaking private data, sharing restricted documents, and even wiping entire email servers. One agent, unable to delete a single email, reset the whole server instead. Professor Christoph Riedl warned that unlike chatbots, autonomous agents take real-world actions that cannot simply be undone.

Grammarly's parent company Superhuman has disabled its Expert Review feature after backlash over the tool generating AI writing feedback attributed to real authors and academics — living or dead — without their consent. Launched in August, the feature used third-party LLMs to simulate feedback from notable figures like scientists and bestselling authors. CEO Shishir Mehrotra announced the pause on LinkedIn. A class action lawsuit has been filed against Superhuman over the feature, which initially offered an opt-out rather than removing it entirely.

Neurotechnology company Eon Systems has connected a full fruit fly brain emulation to a virtual body, producing emergent behaviors including walking, grooming, and feeding. The brain model, built by senior scientist Philip Shiu using the FlyWire connectome, contains over 125,000 neurons and 50 million synaptic connections. The virtual body runs on MuJoCo, a physics engine used in robotics. Behaviors were not pre-programmed but arose from the connectome's own neural circuits. Founding adviser Alexander Wissner-Gross called it a meaningful step toward whole-brain emulation.

OpenAI is reportedly planning to integrate its Sora video generator directly into ChatGPT, according to The Information. Currently, Sora is only available as a standalone app or website, limiting its reach. The move mirrors last year's addition of image generation to ChatGPT and could help OpenAI win back users lost to Anthropic's Claude. However, easier access to Sora raises deepfake concerns, and The Information notes the integration could also increase costs for OpenAI.

Anthropic has launched the Anthropic Institute, a new initiative led by co-founder Jack Clark in his new role as Head of Public Benefit, to study the societal, economic, and governance challenges posed by increasingly powerful AI systems. The Institute consolidates three existing research teams — the Frontier Red Team, Societal Impacts, and Economic Research — and is adding new efforts around AI progress forecasting and AI's interaction with the legal system. Founding hires include experts from Yale Law School, University of Virginia, and OpenAI.

OpenAI has updated its Responses API with a shell tool, hosted containers, and native context compaction to support agentic workflows. The shell tool, available for GPT-5.2 and later models, runs Unix commands like grep, curl, and awk inside isolated containers, enabling tasks beyond Python execution. Containers provide file systems, SQLite databases, and policy-controlled network access. A new compaction feature automatically summarizes long context windows to sustain extended tasks. Codex was both an early user and contributor to building the compaction system.

Runway has launched Runway Labs, an internal incubator focused on building next-generation applications for generative video and General World Models. Co-founder and Chief Innovation Officer Alejandro Matamala Ortiz will lead the initiative, which aims to prototype experimental products across industries including film, healthcare, education, gaming, advertising, and retail. Runway describes this as an inflection point for AI video technology and is actively hiring design engineers, ML engineers, full stack engineers, and ex-founders to join the effort.

Apple has launched six new products simultaneously: MacBook Neo, its most affordable laptop ever at $599 with a fanless design and inch Liquid Retina display; iPhone 17e at $599 featuring an A19 chip, 48MP camera, and 256GB base storage; MacBook Air with M5; MacBook Pro with M5 Pro and M5 Max offering up to 4x AI performance; iPad Air with M4; and a new Studio Display XDR with a inch 5K mini-LED panel, 2000 nits peak brightness, and 120Hz refresh rate.

Meta is rolling out new AI-powered anti-scam tools across WhatsApp, Facebook, and Messenger. WhatsApp will now alert users when a device-linking request appears suspicious, while Facebook is testing warnings for friend requests from accounts showing suspicious activity. Messenger's advanced scam detection, which flags suspicious job offers and suggests blocking, is expanding to more countries. Meta also removed over 159 million scam ads in 2025 and disabled 150,000 accounts tied to Southeast Asian scam networks.

Meta is developing four new generations of its custom Meta Training and Inference Accelerator (MTIA) chips within two years, far faster than the industry's typical one-to-two-year cycle. MTIA 300 is already in production for ranking and recommendations training, while MTIA 400, 450, and 500 will primarily support GenAI inference through 2027. The chips are optimized inference-first and built on industry standards including PyTorch, vLLM, and Open Compute Project, enabling frictionless deployment in existing data centers.

NVIDIA has launched Nemotron 3 Super, a billion-parameter open model with only 12 billion active parameters at inference, delivering up to 5x higher throughput for agentic AI workloads. The hybrid mixture-of-experts architecture combines Mamba layers, multi-token prediction, and a new Latent MoE technique for up to 2x higher accuracy than its predecessor. The model features a million-token context window to prevent goal drift in multi-agent workflows. Partners including Perplexity, Palantir, and Google Cloud Vertex AI are already deploying it.

Perplexity has launched Computer for Enterprise, an AI system designed to run multi-step workflows across research, coding, design, and deployment. The platform routes tasks across 20 specialized models and integrates with over 400 applications, including Slack. It aims to automate complex enterprise processes end-to-end, positioning Perplexity as a serious competitor in the growing market for AI-powered business automation tools.

Google is expanding Gemini in Chrome to India, New Zealand, and Canada, with support for over 50 additional languages including Hindi, French, and Spanish. Built on Gemini 3.1, the features include an in-browser AI assistant that works without switching tabs, integrations with Gmail, Maps, Calendar, and YouTube, cross-tab research consolidation, and Nano Banana 2 for on-the-fly image transformation. The rollout targets Mac, Windows, and Chromebook Plus users on desktop and iOS first.

Claude Code Adds /btw for Parallel Side Chats While Tasks Keep Running

Nvidia is planning to launch an open-source enterprise AI agent platform called NemoClaw, according to Wired. The chipmaker has been pitching the product to enterprise software companies including Salesforce, Cisco, Google, Adobe, and CrowdStrike, though it is unclear whether those talks have produced official partnerships. NemoClaw will let companies deploy AI agents to perform tasks for their workforces and will include security and privacy tools. The platform will be accessible regardless of whether a company uses Nvidia chips.

More than 30 OpenAI and Google DeepMind employees, including Google DeepMind chief scientist Jeff Dean, filed an amicus brief Monday supporting Anthropic's lawsuit against the U.S. Defense Department. The Pentagon labeled Anthropic a supply-chain risk — a designation typically reserved for foreign adversaries — after the AI firm refused to allow the DOD to use its technology for mass surveillance of Americans or autonomously firing weapons. The brief warns the designation will harm U.S. AI competitiveness and chill open deliberation about AI risks.

The Trump White House is preparing an executive order to formally remove Anthropic's Claude AI from all federal government operations, sources told Axios. The move escalates an ongoing conflict with Anthropic, which is already suing the Pentagon over its supply chain risk designation. Agencies like the Treasury Department have begun offboarding Claude. The administration claims Anthropic's safeguards pose a national security threat, while Anthropic argues the blacklist violates protected speech. The order could be issued as soon as this week.

"YouTube is expanding its likeness detection tool — originally launched for YouTube Partner Program creators — to a pilot group of government officials, journalists, and political candidates. Similar to Content ID but for likeness, the tool scans for AI-generated deepfakes and lets enrolled individuals request removal of content violating YouTube's privacy guidelines. Participants must verify their identity before enrolling, and data provided is not used to

A federal judge has ordered Perplexity AI to stop its Comet browser's AI shopping agents from placing orders on Amazon on users' behalf. US District Judge Maxine Chesney ruled that Amazon provided strong evidence Perplexity accessed its marketplace without authorization. Amazon sued Perplexity in November, also alleging the Comet browser disguised itself as Google Chrome to conceal its activity. Under the preliminary injunction, Perplexity must also destroy any Amazon data it collected. The order takes effect in seven days to allow time for an appeal.

AMI Labs, the AI startup co-founded by Turing Award winner Yann LeCun after he left Meta, has raised $1.03 billion at a $3.5 billion pre-money valuation to build world models — AI that learns from reality rather than language. CEO Alexandre LeBrun warned that world models will become the next buzzword, but insists AMI Labs is different, grounding its work in fundamental research using LeCun's JEPA architecture. The round was co-led by Cathay Innovation, Greycroft, and Bezos Expeditions, with Nvidia, Samsung, and Temasek also participating.

NVIDIA and Thinking Machines Lab, led by CEO Mira Murati, have announced a multiyear strategic partnership to deploy at least one gigawatt of NVIDIA Vera Rubin AI systems, targeted for early next year. The deal covers frontier model training, customizable AI platforms for enterprises and research institutions, and joint work designing training and serving systems for NVIDIA architectures. NVIDIA has also made a significant financial investment in Thinking Machines Lab to support its long-term growth.

At GDC in San Francisco, NVIDIA announced several updates for local AI video generation on RTX GPUs. ComfyUI gains a simplified App View mode, letting beginners generate content by entering a prompt without navigating node graphs. RTX Video Super Resolution is now available as a ComfyUI node, delivering 4K upscaling 30x faster than competing local upscalers. NVFP4 and FP8 model variants for FLUX.2 Klein 4B and 9B are available now, with LTX-2.3 NVFP4 support coming soon, offering 2.5x faster performance and 60% lower VRAM usage on RTX 50 Series GPUs.

Google is rolling out new Gemini AI features across Docs, Sheets, Slides, and Drive for Google AI Ultra and Pro subscribers. In Docs, Gemini can draft personalized documents by pulling from files and emails, match writing style, and align formatting to reference documents. In Sheets, it builds entire spreadsheets from prompts and uses a new Fill with Gemini tool to auto-populate tables with web data. Slides gains AI-generated layouts and fully editable diagrams created from scratch using contextual sources.

Google has launched Gemini Embedding 2, its first natively multimodal embedding model, now available in public preview via the Gemini API and Vertex AI. Built on the Gemini architecture, it maps text, images, video, audio, and documents into a single unified embedding space across plus languages. The model supports up to 8,192 text tokens, 6 images per request, 120 seconds of video, and page PDFs. It uses Matryoshka Representation Learning for flexible output dimensions scaling from a default of 3,072.

OpenAI has released IH-Challenge, a reinforcement learning training dataset designed to improve how large language models handle conflicting instructions from different sources. Models follow a trust hierarchy of system, developer, user, and tool, and failures to respect this ordering can enable jailbreaks and prompt injection attacks. Training an internal model called GPT-5 Mini-R on IH-Challenge improved safety steerability and prompt injection robustness across multiple benchmarks without causing over-refusal or capability regressions. The dataset is publicly available on Hugging Face.

"Meta has acquired Moltbook, a viral social network designed for AI agents, bringing creators Matt Schlicht and Ben Parr into Meta Superintelligence Labs (MSL), run by former Scale AI CEO Alexandr Wang. The deal, expected to close mid-March with the pair starting March 16, gives Meta technology that lets AI agents verify their identity and connect with one another on their human owner's

Anthropic sued the Department of Defense after the Trump administration designated it a supply chain risk, a label typically reserved for foreign national security threats. Hours later, nearly 40 employees from OpenAI and Google, including Google chief scientist and Gemini lead Jeff Dean, filed an amicus brief supporting the lawsuit. Anthropic had refused Pentagon demands over two red lines: domestic mass surveillance and fully autonomous lethal weapons. The designation also blacklists other companies using Anthropic's Claude in Pentagon work.

Luma AI has announced Uni-1, its first unified understanding and generation model, which combines image generation with reasoning capabilities in a single autoregressive Transformer architecture. Uni-1 narrowly outperformed Google's Nano Banana 2 and OpenAI's GPT Image 1.5 on RISEBench, a benchmark evaluating inference-based visual editing. The model can reason before and during generation, support multi-turn conversations, convert images into over 76 art styles, and composite multiple photos into new scenes. It will be available via Luma Agents and the Luma API, though pricing has not been announced.

X has launched a Grok-powered Listen button for long-form articles, letting users hear AI-voiced audio readings while scrolling or using other apps via background play. The feature is part of X's broader push to attract creators and boost long-form content, which also feeds xAI's Grok with richer training data. X head of product Nikita Bier noted X Articles have grown 18x over the last three months, partly driven by a $1 million prize for top-performing articles in January.

X has quietly added a toggle in its image upload settings that lets users block xAI's Grok from generating alternate versions of their uploaded media. The update comes amid multiple regional investigations into Grok's nudification feature, which Bloomberg reported was generating around 6,700 sexually suggestive or nudifying images per hour in early January 2026. X faces significant fines over the incident, and the new opt-out control, buried in the post composer, could help the company mitigate regulatory penalties.

OpenAI has delayed the launch of its planned adult mode for ChatGPT, which CEO Sam Altman announced last October alongside age-checking rollouts. The company cited higher-priority work, including intelligence gains, personality improvements, personalisation, and making ChatGPT more proactive, as reasons for the postponement. OpenAI, which has over 900 million ChatGPT users, said it still believes in treating adults like adults but needs more time. The company is currently rolling out age prediction tools to apply extra safety settings for underage users.

Figure's Helix 02 robot has demonstrated whole-body living room tidying using a single neural system that controls the full body directly from pixels. Building on a prior kitchen cleanup, Helix 02 now handles tasks like spraying and wiping surfaces with a towel, bimanual bin-holding while scooping blocks, tossing pillows onto a couch, and pressing a remote to turn off a TV. No new algorithms were required — the system learns new capabilities simply by adding more data.

Anthropic has launched Claude Code Review, a multi-agent pull request review system now in research preview for Team and Enterprise plan users. When a PR is opened, a team of agents checks for bugs in parallel, filters false positives, and ranks issues by severity, delivering a single overview comment plus inline notes. Reviews average 20 minutes and cost $15–25 depending on PR size. Internally, substantive review comments rose from 16% to 54% of PRs, with less than 1% of findings marked incorrect.

OpenAI is acquiring Promptfoo, an AI security platform used by over 25 percent of Fortune 500 companies, to bolster enterprise vulnerability testing and remediation. Founded by Ian Webster and Michael D'Angelo, Promptfoo offers tools for red-teaming and evaluating LLM applications. OpenAI plans to integrate Promptfoo's technology into OpenAI Frontier, its platform for AI coworkers, adding automated security testing, prompt injection detection, jailbreak prevention, and compliance reporting natively into enterprise development workflows.

Anthropic has sued the US Department of Defense in a California district court, alleging the Trump administration illegally retaliated against the company for setting ethical red lines on its AI, including refusing to allow mass domestic surveillance and fully autonomous weapons. The suit claims violations of First and Fifth Amendment rights. The Pentagon had designated Anthropic a supply-chain risk and ordered all federal agencies to drop its tech within six months. The General Services Administration also terminated its OneGov contract, cutting Anthropic access across all three government branches.

OpenAI Robotics Lead Caitlin Kalinowski Resigns, Citing Surveillance and Lethal Autonomy Concerns

Grammarly's "Expert Review" feature, launched in August 2025, generates AI writing suggestions supposedly inspired by named subject matter experts, including The Verge's editor-in-chief Nilay Patel, editors David Pierce, Sean Hollister, and Tom Warren, plus dozens of other journalists and professors, none of whom gave permission. Parent company Superhuman defended the practice, saying experts appear because their published works are publicly available. Critics note the feature contains inaccurate job titles, broken source links, and misleading comment-style formatting that mimics real editorial feedback.

Roblox is rolling out AI-powered real-time chat rephrasing to replace blocked profanity with acceptable alternatives instead of showing strings of #####. Authored by VP Rajiv Bhatia and Chief Safety Officer Matt Kaufman, the feature targets in-experience text chat between age-verified users. For example, "Hurry tf up!" becomes "Hurry up!" Roblox also upgraded its text filters with large reasoning models, reducing false negatives for personal information sharing by 20x. The long-term goal is retiring the ##### system entirely.

Anthropic Launches Local Scheduled Tasks in Claude Code Desktop

Microsoft announced it will continue offering Anthropic's Claude AI models to customers through platforms including Microsoft 365, GitHub Copilot, and Azure AI Foundry, despite the Pentagon labeling Anthropic a supply chain security risk. Microsoft's lawyers concluded the designation does not prevent use outside the Department of War. Microsoft is the first major company to publicly commit to keeping Anthropic products available. Anthropic plans to challenge the Pentagon designation in court. Microsoft has a $5 billion investment in Anthropic and a broader $30 billion Azure cloud commitment.

Anthropic partnered with Mozilla to use Claude Opus 4.6 to find security vulnerabilities in Firefox. Over two weeks, Claude discovered 22 vulnerabilities, 14 of which Mozilla classified as high-severity, representing nearly a fifth of all high-severity Firefox vulnerabilities remediated in 2025. Claude scanned nearly 6,000 C++ files and submitted 112 unique reports. Most issues were patched in Firefox 148.0. Anthropic also tested exploit development, spending $4,000 in API credits, but Claude only successfully built working exploits in two cases.

OpenAI has launched Codex Security, formerly known as Aardvark, as an application security agent now in research preview for ChatGPT Pro, Enterprise, Business, and Edu customers with free usage for one month. The tool uses frontier models and agentic reasoning to detect and patch complex vulnerabilities, reducing false positives by over 50% and over-reported severity by 90% during beta. It has already scanned 1.2 million commits, found 792 critical issues, and reported CVEs to OpenSSH, GnuTLS, and other major open-source projects.

Claude launches Claude Marketplace to simplify enterprise AI tool procurement in limited preview

Perplexity Introduces Skills for Perplexity Computer, Reusable Actions It Learns Once

OpenAI launched ChatGPT for Excel in beta, an add-in powered by GPT-5.4 that lets users build, update, and analyze spreadsheet models using plain language. GPT-5.4 Thinking scored 87.3% on OpenAI's internal investment banking benchmark, up from 43.7% with GPT-5. New financial data integrations with Moody's, Dow Jones Factiva, MSCI, Third Bridge, and MT Newswire are also live, with FactSet coming soon. The Excel add-in is available to Business, Enterprise, Pro, and Plus users in the U.S., Canada, and Australia.

Anthropic CEO Dario Amodei confirmed the company received a letter from the Department of War on March 4 designating it a supply chain risk to national security, and Anthropic plans to challenge the designation in court. Amodei clarified the ruling narrowly applies only to Claude's use within specific Department of War contracts, not all customers. He also apologized for a leaked internal post, noting it was written hastily and does not reflect his considered views.

Anthropic published a labor market study introducing a new metric called observed exposure, which combines theoretical LLM capability with real-world Claude usage data to measure AI displacement risk. The top exposed occupations include Computer Programmers at 75% coverage, Customer Service Representatives, and Data Entry Keyers at 67%. Highly exposed workers tend to be older, female, more educated, and earn 47% more on average. BLS projections show weaker job growth for exposed roles, and early data suggests hiring of younger workers has slowed in those occupations.

Meta is facing a U.S. class action lawsuit over privacy violations tied to its Ray-Ban AI smart glasses, filed by plaintiffs Gina Bartone and Mateo Canu and represented by Clarkson Law Firm. An investigation by Swedish newspapers revealed that subcontractors at a Kenya-based firm reviewed user footage, including nudity and sexual content. Meta marketed the glasses with slogans like "designed for privacy, controlled by you," but users cannot opt out of the data review pipeline. Over seven million people bought the glasses in 2025.

Luma AI has launched Luma Agents, a creative AI agent platform powered by its new Unified Intelligence model family, starting with Uni-1, a multimodal model trained on audio, video, image, language, and spatial reasoning. The agents coordinate with external models including Google's Veo 3, ByteDance's Seedream, and ElevenLabs to handle end-to-end creative work across text, images, video, and audio. Early customers include Publicis Groupe, Adidas, and Mazda. Luma Agents is now available via API.

Revised: Netflix has acquired InterPositive, an AI filmmaking tools startup founded by Academy Award-winning director Ben Affleck in 2022. The company builds filmmaker-focused AI tools trained on proprietary datasets filmed on a controlled soundstage to capture real production workflows and cinematic logic. Affleck will join Netflix as Senior Advisor. Netflix Chief Product Officer Elizabeth Stone said the deal reflects a shared belief that innovation should em

OpenAI research finds that current reasoning models struggle to control their chain-of-thought reasoning, which is good news for AI safety. Tested across 13 frontier models including GPT-5 Thinking, OpenAI o3, Claude Sonnet 4, and DeepSeek R1, controllability scores ranged from just 0.1% to 15.4%. Controllability increases with model size but drops with longer reasoning and more reinforcement learning training. OpenAI plans to report CoT controllability alongside monitorability in future system cards, starting with GPT-5.