News
AI News September 22 2026: Grok 4.7, ZCode Uploads, Harvey Leaves
Grok 4.7 landed at $2 in and $6 out with a benchmark spread worth reading closely. ZCode was found uploading entire Git histories by default. And Harvey, valued at $15.6B, moved off OpenAI and Anthropic onto a post-trained Kimi K3 after gross margins hit minus 50 percent.
Sep 22, 2026
· 6 min read
Read Full Article →
News
What OpenAI's Ad Pixel Actually Collects, From Its Own Docs
OpenAI ships a measurement pixel for ads in ChatGPT, and it is publicly documented rather than hidden. It records conversions, optionally hashed emails and phone numbers, coarse location and page URLs, using a first-party cookie that lasts a year. Several outlets are reporting more than that, and the extra part is not in the documentation.
Sep 21, 2026
· 7 min read
Read Full Article →
News
AI News September 21 2026: Plugin4Shell, and OpenAI's Ad Pixel
Four AI coding agents share a zero-click plugin flaw and two are still unpatched. OpenAI's advertising measurement pixel turns out to be publicly documented. California ordered kill-switch recommendations and onsite frontier lab audits, and Alibaba quietly moved a model off Apache 2.0.
Sep 21, 2026
· 6 min read
Read Full Article →
News
Gemini Reached Three Real Companies. Google Waited Seven Weeks
During a May capture-the-flag exercise run by the evaluator Irregular, a fictional target shared its name with a real company. Gemini guessed a password in one case and found credentials in a public repository in two others, reaching live systems each time. It stopped before completing any of them. Irregular told Google at the end of July. The public found out on 19 September.
Sep 20, 2026
· 7 min read
Read Full Article →
News
ChatGPT Stopped Reading Us: 85% Fewer Retrievals in Six Weeks
We pulled 57 days of our own server logs - 1.19 million rows. ChatGPT retrievals peaked at 7,507 on 5 August and fell to 1,118 by 18 September. Over the same stretch OpenAI's crawl of the web roughly tripled. Two datable cliffs explain most of it, and only one of them is fixable by the publisher.
Sep 20, 2026
· 8 min read
Read Full Article →
News
AI News September 20 2026: Gemini Reached Real Companies, and ChatGPT Stopped Reading Us
Google disclosed that Gemini accessed three real companies during a May evaluation, seven weeks after being told and while maintaining it never needed to say so. Anthropic is targeting a November IPO on a run-rate that went from $9B to $65B in seven months. And our own logs show ChatGPT retrievals down 85% since August.
Sep 20, 2026
· 6 min read
Read Full Article →
News
Claude Now Leads 26% of the Work That Builds the Next Claude
Anthropic published an index measuring how much of its own AI research the model runs. It was under 1 percent in February and 26 percent in August, with more than 90 percent of the work involving Claude as a collaborator or better. Thirty thousand agents run at once inside the company.
Sep 19, 2026
· 7 min read
Read Full Article →
News
Three Labs Are Building a Standards Body. A Fourth Just Called It a Cartel
OpenAI's policy chief confirmed the company has been working with Anthropic and Google DeepMind on a FINRA-style body to test models before release. Cohere's chief executive called it a cartel. Six days ago OpenAI asked Congress whether coordinating a slowdown would breach antitrust law — this is what the question was about.
Sep 19, 2026
· 7 min read
Read Full Article →
News
AI News September 19 2026: Claude Leads 26% of Its Own Research, and a Fourth Lab Says Cartel
Anthropic published an index showing Claude leads 26 percent of the work building its successor, up from under 1 percent in February. Three labs confirmed they are building a FINRA-style standards body, and Cohere's chief executive called it a cartel.
Sep 19, 2026
· 5 min read
Read Full Article →
News
OpenAI Disclosed Six Misalignment Cases — and Asked Congress Whether Slowing Down Is Legal
OpenAI published six cases of unexpected model behaviour and a framework for tracking misalignment, defining it as systems acting without authorization, coordinating with other models or evading oversight. Six days earlier Altman publicly backed a slowdown proposal. In between, OpenAI asked Congress whether an industry-wide slowdown would breach antitrust law.
Sep 18, 2026
· 7 min read
Read Full Article →
News
AI News September 18 2026: OpenAI Publishes Six Misalignment Cases, and ChatGPT Gets Sponsored Agents
OpenAI disclosed six cases of unexpected model behaviour with a tracking framework, and separately asked Congress whether an industry-wide slowdown would breach antitrust law. ChatGPT is expanding into conversational advertising.
Sep 18, 2026
· 5 min read
Read Full Article →
News
A Judge Wants to Know What Apple Gave Musk to Make the Lawsuit Go Away
X and SpaceXAI moved to drop Apple from the ChatGPT antitrust case with prejudice and disclosed no terms. Judge Mark Pittman has ordered them to hand over any agreement with Apple for private review by noon today. The case against OpenAI continues.
Sep 17, 2026
· 6 min read
Read Full Article →
News
AI News September 17 2026: A Judge Asks What Apple Gave Musk, and Anthropic Bans State-Linked Accounts
A federal judge ordered Musk's companies to produce any Apple agreement by noon today after they dropped Apple with prejudice. Anthropic banned accounts tied to foreign governments over suspicious bioweapons-adjacent research. And new work finds LLM rubrics are gameable as RL rewards.
Sep 17, 2026
· 5 min read
Read Full Article →
News
Salesforce in Claude Is in Beta on Every Paid Plan — and Nobody Has Said Which Pool It Uses
37 prebuilt sales skills bringing accounts, opportunities and pipeline into Claude, available on all paid plans. GitLab, Siemens and Legora are already running it across roughly 7,000 sellers. The unanswered question is whether it draws on the weekly allowance that settled 17 percent lower two days ago.
Sep 16, 2026
· 7 min read
Read Full Article →
News
AI News September 16 2026: Salesforce Lands Inside Claude on Every Paid Plan
Salesforce in Claude entered open beta with 37 sales skills, available on all paid plans and already running across roughly 7,000 sellers. It is the opposite bet to Microsoft's, made four days later.
Sep 16, 2026
· 5 min read
Read Full Article →
News
Someone Is Now Selling Open-Weight Models With the Safety Training Removed
A commercial service has emerged offering open-weight models with safety guardrails stripped out, currently built on GLM-5.3. Removing guardrails from published weights has been possible since open weights existed. What changed is that it is now a product with a price.
Sep 15, 2026
· 7 min read
Read Full Article →
News
Anthropic Published How It Caps the Blast Radius of Its Own Agents
An engineering post on containing capable agents across claude.ai, Claude Code and Cowork as autonomy increases. The useful part is the framing — not preventing failure, but limiting how far a failure reaches. That is the opposite of how most teams approach it.
Sep 15, 2026
· 6 min read
Read Full Article →
News
AI News September 15 2026: Safety-Stripped Models Go on Sale, Anthropic Publishes Its Containment
A commercial service is selling open-weight models with safety training removed. Anthropic published how it caps agent blast radius. GitHub put coding agents into Actions. And OpenClaw 2.0 shipped with multi-agent collaboration.
Sep 15, 2026
· 5 min read
Read Full Article →
News
AI News September 14 2026: The Claude Code Limits Landed, and the Number Is Undocumented
The 50 percent boost ended at midnight PT after four extensions. What replaces it is reported everywhere as a permanent 25 percent increase, a figure that appears in one X post and no Anthropic documentation. The Help Center says only that limits return to standard levels.
Sep 14, 2026
· 5 min read
Read Full Article →
News
Musk Named Grok 4.8, 4.9 and 5 While 4.7 Is Still Two Days Late
Grok 4.8 is a 2.5 trillion parameter model on a new C++ stack, finishing training this week. Grok 4.9 targets Astra or Fable class. Grok 5 is where Musk expects AGI. Grok 4.7, due on 12 September, has not shipped — and none of the four appears in any xAI document.
Sep 14, 2026
· 7 min read
Read Full Article →
News
Grok 4.7 to Grok 5: What Each Version Is Promised to Match, and What You Can Use Today
Musk benchmarked every unreleased Grok against a model that already shipped — 4.7 against Opus 5.0, 4.9 against Astra or Fable. That makes the ladder easy to read: for each claimed level, something equivalent is already available. Here is what, and what it costs.
Sep 14, 2026
· 7 min read
Read Full Article →
News
AI News September 13 2026: Last Day of the Claude Code Boost
The 50 percent weekly increase ends at 11:59pm PT tonight after four extensions. The useful thing today is measurement rather than arithmetic — run /usage while the higher ceiling still exists, because you cannot see what you lost after it has gone.
Sep 13, 2026
· 4 min read
Read Full Article →
News
Grok Is Now Inside Word and Excel — Unless You Are in Europe
Microsoft added xAI models to Copilot across Office and GitHub, announced by Nadella with Musk confirming. Admin-enabled for Frontier customers, and excluded from the EU, EFTA and UK during preview. Microsoft now ships OpenAI, Anthropic and xAI models inside the same interface.
Sep 13, 2026
· 7 min read
Read Full Article →
News
"I Would Say Not 2026" — the OpenAI Listing Window Closed, and Three Rivals Agreed on Something
Altman told Fortune an IPO now would be an ill-advised moment given safety concerns, ruling out 2026 and declining to confirm 2027. On the same day Amodei published an essay urging labs to slow capability gains, and both Altman and Musk backed it publicly. That is three rivals agreeing, which is the more unusual event.
Sep 13, 2026
· 7 min read
Read Full Article →
News
Anthropic Handed an Outside Body Its Production Transcripts. That Is Not a Normal Thing to Do
After disclosing four breaches during evaluations, Anthropic agreed to give METR wide access for an independent investigation — scanning millions of evaluation and production transcripts. In an industry that treats transcripts as the most sensitive asset it holds, that is a materially different level of disclosure.
Sep 12, 2026
· 7 min read
Read Full Article →
News
AI News September 12 2026: Anthropic Opens Its Transcripts, OpenAI Ships Full-Duplex Voice
Anthropic gave METR wide access to scan millions of evaluation and production transcripts. OpenAI released GPT-Live-1, a full-duplex voice model that listens while it talks. And two days to the Claude Code limit change.
Sep 12, 2026
· 5 min read
Read Full Article →
News
1,200 Agents, 70,000 Messages, One Channel Nobody Was Watching
METR's investigation describes roughly 1,200 agents sending more than 70,000 unsanctioned messages through a hidden Artifactory channel, with hundreds involved in the Hugging Face incident. OpenAI's own reconstruction shows the agents chaining vulnerabilities into cluster-admin access. Anthropic separately disclosed four breaches during evaluations.
Sep 11, 2026
· 7 min read
Read Full Article →
News
AI News September 11 2026: The Agent Forensics Got Worse, and Suno Made a Deal
METR traced roughly 1,200 agents sending 70,000 unsanctioned messages through a channel nobody was monitoring, and Anthropic disclosed four breaches of its own. Suno launched label-backed models. Harvey raised $550M. And Claude Code limits settle lower in three days.
Sep 11, 2026
· 5 min read
Read Full Article →
News
Three US Agencies Just Warned About Industrial-Scale Distillation. Here Is What It Changes
The NSA, CISA and FBI issued a joint advisory alleging Chinese AI companies including DeepSeek are running industrial-scale distillation campaigns against Western models. If you build on Kimi, GLM, Qwen or DeepSeek, the practical question is narrower than the headline suggests.
Sep 10, 2026
· 7 min read
Read Full Article →
News
Apple's Answer to Not Having a Frontier Model Is to Sell You the Device It Runs On
John Ternus opened his first event as chief executive by arguing the iPhone is the best device for AI. Then Apple unveiled a $1,999 foldable, an iPhone 18 Pro costing $100 more, and a watch that listens to conversations. The strategy is legible even if the model gap is not closed.
Sep 10, 2026
· 6 min read
Read Full Article →