AI News
AI news that matters. Updated daily.
No stories match your filters.
Fake Privacy Filter Model on Hugging Face Confirmed as Credential-Stealing Malware
A model called Open-OSS/privacy-filter on Hugging Face - which presents itself as an OpenAI privacy filtering tool - has been confirmed as a multi-stage credential-stealing malware package. Developers who downloaded and ran it should treat their machines as compromised.
Claude Sonnet 4.6 Resetting Mid-Response and Charging Tokens for Failed Answers
Multiple users are reporting that Claude Sonnet 4.6 is cutting off mid-response, resetting the conversation to its starting state, and consuming tokens without producing any output. The issue surfaced in user reports on May 7, 2026.
97% Less Context Fixed Claude's Hallucinations Across 18 Real Codebases
80,000 tokens of source code per query was producing one consistent result in a developer's codebase: Claude confidently suggesting functions that didn't exist anywhere in the project.
Anthropic's Colossus Compute Deal Puts Its Public Benefit Status Under a Microscope
The Anthropic-Colossus compute deal is creating a trust problem the company may not have fully anticipated.
OpenAI Adds Voice Intelligence Features to Its Developer API
OpenAI just added a set of voice intelligence features to its API, giving developers tools to build voice-powered products across customer service, education, and content creation.
50 LLMs Took 45 Psychology Tests. The Results Aren't Personality.
Researchers gave 45 psychological questionnaires - the kind used to measure personality traits, cognitive styles, and mental health markers in humans - to 50 different large language models. The conclusion: what these tests appear to capture in AI has nothing to do with what they measure in people.
Claude Pro's 'No Limits' Marketing May Violate EU Consumer Protection Law
"Without hitting limits on the Pro plan." "No more interruptions mid-task." Those are direct quotes from Anthropic's Claude Pro pricing page. EU subscribers who have hit those limits mid-task have a legitimate consumer protection question.
Google Quietly Installed a 4GB Gemini Model in Chrome - Here's the Trade-Off
What happens when a browser silently downloads a 4GB AI model onto your computer without ever asking permission? Chrome users recently discovered that Google had quietly bundled a local Gemini AI model into Chrome - a 4-gigabyte file that arrived without a notification, a consent prompt, or an opt-in.
Opus 4.7 Classifiers Are Silently Terminating Biology and Cybersecurity Conversations
Three weeks into Opus 4.7's release, a specific group of professional users - security researchers, biology educators, virology professionals - is running into the same wall: conversations about their core subject matter are being terminated without explanation.
OpenAI Adds 'Trusted Contact' Safety Feature to ChatGPT for Crisis Situations
OpenAI is adding a feature called Trusted Contact to ChatGPT that lets users designate a person - a friend, family member, or therapist - to be notified when a conversation suggests they may be at risk of self-harm. The company announced the feature via TechCrunch on May 7, 2026.
Perplexity Personal Computer Opens to All Mac Users
Perplexity just opened its Personal Computer feature to all Mac users, dropping the access restrictions that had previously limited it. The product brings AI agents - software that takes actions on your computer rather than just answering questions in a chat window - to the Mac desktop.
Murati's Deposition Fills In New Details on Altman's 2023 Firing
The week before Thanksgiving 2023 was the most chaotic moment in AI corporate history: Sam Altman was fired as OpenAI's CEO on November 17th, then reinstated five days later after nearly the entire company threatened to resign. The board's official explanation was that Altman was "not consistently candid in his communications with the board" - a statement thin enough to generate years of speculation about what actually happened behind closed doors.
Two South African Officials Suspended After AI Hallucinations Found in Government Documents
Two officials at South Africa's Department of Home Affairs were suspended after AI-generated hallucinations were discovered in official government documents. The suspensions, reported in May 2026, are among the first publicly documented cases of government employees facing formal disciplinary action for submitting AI output containing factual errors.
A Developer Built a Complete Physics Pinball Game Using Only Claude
What happens when someone with no game engine experience sets out to build a full physics pinball game using Claude as their only coding partner? One developer just published a detailed postmortem answering that question, and the honest account of what worked and what didn't is more useful than most AI coding tutorials.
Musk's OpenAI Lawsuit Forces a Public Reckoning on AI Safety Governance
Three years after Elon Musk co-founded OpenAI with Sam Altman as a nonprofit devoted to building safe AI, he's suing the company for abandoning that mission. The lawsuit has wound through courts for over a year, and it's now doing something beyond the legal dispute itself: forcing a public audit of whether OpenAI's safety commitments have been substantive or performative.
Mozilla Found 271 Real Vulnerabilities in Firefox Using AI, Almost No False Alarms
271 vulnerabilities. That's how many real security flaws Mythos - an AI-powered code analysis tool - surfaced in Mozilla's Firefox codebase. More striking than the number: Mozilla says the results came back with "almost no false positives."
Anthropic Removes Email Reading from Claude's Gmail Connector
Anthropic removed the gmailreadmessage capability from Claude's Gmail connector, breaking workflows that depended on the model reading incoming emails.
ChatGPT Gets 'Trusted Contact' Feature to Flag Self-Harm Conversations to Loved Ones
OpenAI is adding a safety net to ChatGPT that lets adult users designate a "Trusted Contact" - a friend, family member, or caregiver who gets notified if the AI detects concerning conversations about self-harm or suicide.
AWS Builds Payment Infrastructure for Autonomous AI Agents
What happens when an AI agent needs to pay for something it just ordered on your behalf? AWS just shipped an answer.
Anthropic CEO: 80x Q1 Growth Explains Claude's Compute Problems
80x. That's how much Anthropic grew in the first quarter of 2026 on an annualized basis, according to CEO Dario Amodei speaking at the company's developer conference in San Francisco.
Claude Code's CLI Has No Scroll History - and Months of Reports Haven't Fixed It
Scroll up in Claude Code's CLI (command-line interface - the version that runs in your terminal) and you'll see nothing. Not the previous output, not an error message - just blank space where your conversation history should be.
ChatGPT in Chinese Has a Soothing Phrase Problem: 'I'll Catch You Steadily'
The ChatGPT chatbot behaves noticeably differently depending on what language you're speaking - and Chinese users have started documenting the gap.
Anthropic's Mythos Found High-Severity Firefox Bugs That Years of Auditing Missed
Mozilla's security team has a problem most software companies would envy: Firefox is one of the most-audited codebases in the world, with thousands of contributors, continuous automated testing, and decades of hardening. Yet Anthropic's Mythos - an AI-powered security research tool - walked in and found a wealth of high-severity bugs that all of that existing process had missed.
EU Member States Vote to Ease AI Act Restrictions Before Full Rollout
Last year, the EU AI Act looked like the floor, not the ceiling - a framework everyone expected to grow stricter over time. Now EU member states have approved a provisional deal to roll back parts of it, reversing that assumption before the law had fully phased in.
Anthropic Expands Claude Usage Limits for Enterprise Customers
Anthropic struck a compute deal with SpaceX this week and paired it with changes that directly affect paying customers. According to AI Business, the company expanded usage limits and reduced the access restrictions that have frustrated enterprise Claude users.
Anthropic Secures SpaceX Colossus 1 Access as Valuation Reportedly Hits $1.2T
$1.2 trillion. That's the reported valuation Anthropic has now reached - up roughly 80x from its early-stage numbers - as the company secures the hardware needed to compete at the frontier. Reports indicate Anthropic has locked in dedicated access to SpaceX's Colossus 1, one of the largest GPU (graphics processing unit - the chips that power AI training) clusters currently in operation.
AMD's MI350P Brings CDNA 4 AI Performance to Standard PCIe Server Slots
Most of AMD's high-end AI accelerators require SXM slots - the proprietary high-bandwidth sockets found in dedicated AI server platforms that cost significantly more than standard hardware. The new Instinct MI350P doesn't.
Moonshot AI Raises $2B at $20B Valuation as Revenue Hits $200M ARR
$2 billion. That's what investors just put into Moonshot AI, the Chinese startup behind the Kimi chatbot, in a round that values the company at $20 billion.
Claude's Verbal Tics: Why Users Can't Prompt Away 'That's Not Nothing'
"That's not nothing." "That matters." "You're absolutely right." "I genuinely think..."
Spotify Plans to Let Users Import AI-Generated Podcasts From Codex and Claude Code
Spotify is making a direct play for the AI-generated audio market, announcing plans to let users create podcasts using tools like OpenAI's Codex and Claude Code and import them directly to the platform.
Spotify's AI DJ Adds French, German, Italian, and Brazilian Portuguese
Spotify's AI DJ is getting less American. The feature, which generates personalized music mixes narrated by an AI voice that comments on your listening habits and mood, now speaks French, German, Italian, and Brazilian Portuguese. The expansion was announced on May 7, 2026.
Thousands of Vibe-Coded Apps Are Leaking Corporate Data Publicly
What happens when millions of people build software without knowing what authentication means? You get thousands of corporate databases sitting open on the public internet.
A Reggae Band's Viral Hit Was Partly Driven by Unauthorized AI Remixes
Seven years after its release, Stick Figure's "Choice Is Yours" shot up the charts. The band was thrilled - until they looked closer at what was driving it. A significant portion of the song's viral momentum came from unauthorized AI remixes circulating across streaming platforms and social media. The band never approved them. No royalties flowed back to the artists.
DeepSeek Nears $45 Billion Valuation as China's State Fund Moves Into AI
$45 billion. That's the valuation reportedly attached to DeepSeek as China's state-backed "Big Fund" leads investment talks with the Chinese AI lab.
OpenAI Adds Reasoning and Translation to Its Realtime Voice API Models
OpenAI announced new realtime voice models in its API on May 7 that go significantly beyond what its previous voice offering could do. The new models can reason through what they hear, translate between languages, and transcribe speech - all in real time, without routing audio through a separate speech-to-text step.
The Claude Code Comprehension Problem Nobody Talks About
The problem isn't that Claude Code writes bad code. It's that it writes code you don't understand - and you don't notice until something breaks.
Anthropic Details New Training Stage That Makes AI Alignment Actually Generalize
One of the trickier problems in AI development isn't teaching a model to behave well - it's making those behaviors hold up after additional training on top. Anthropic researchers have published details on a new approach called "model spec midtraining" that inserts a dedicated training stage into the pipeline to address this directly.
AI Coding Tools Build Fast. Then You Spend Two Weeks Debugging.
Three days to build something that normally takes three weeks. That's the promise of AI coding assistants, and honestly, it delivers. The initial build phase with tools like Claude Code or Cursor feels close to magic - commits flying, features appearing, momentum compounding on itself.
Non-Coder Built 3 Browser Games with Claude. They Got 25 Million Plays.
25 million plays. That's what three browser games built by someone with zero coding experience three months ago have accumulated.
ChatGPT's New Trusted Contact Feature Notifies Someone If Self-Harm Risk Is Detected
What happens when someone using ChatGPT shows signs of a mental health crisis and there's no one else in the room? OpenAI's answer is Trusted Contact - an optional feature that routes a human into the loop when the model detects serious self-harm concerns.
OpenAI Starts Testing Ads in ChatGPT to Fund Free Access
Ads are coming to ChatGPT. OpenAI announced it's running its first advertising tests inside the product, framing the move as a way to keep free access funded without leaning harder on paid subscriptions.