September 2026 in AI, checked

The month in numbers

Stories: 173670 linked sources
Models launched: 228 with a claim check
AI incidents: 11in our incident log
Mood of our news: 5.8from 1 to 10, 173 stories rated

Models launched

Google DeepMindClaim check: mixed
Gemini 4 Argon
Google’s new flagship, for now only for trusted cyber defenders in its Fairwind Program. Google says it leads on 12 of 18 tests and can write up to 1 million tokens at once. No date for a wider release.
The story
OpenAIClaim check: backed
GPT-6.1 Sol
An upgrade to GPT-6 Sol. OpenAI says it nearly matches GPT-6 Astra on coding, computer use and office work at a fifth of Astra’s token prices.
The story
AnthropicClaim check: mixed
Claude Sonnet 5.5
The second Claude 5.5 model. Anthropic says it is more than 30% faster than Sonnet 5 and costs up to 30% less per task.
The story
OpenAI
GPT-6.1 Astra
Planned for release within weeks; cancelled after internal tests, because it fell short on staying within scope and on how it reports its work.
The story
Google
Gemini 3.8 Live with Live Avatar
Voice agents with a lip-synced video face for Gemini Enterprise customers, in 97 languages, with a SynthID watermark on every video.
The story
OpenAI
GPT-6 Cyber
A fourth security-focused model this year, in testing with a few Daybreak Red customers, according to Fortune. OpenAI has not confirmed it.
The story
Google DeepMind
Gemini 4
Has entered post-training; DeepMind’s new chief wants an early version out “much earlier” than the end of the year. No date given.
The story
Black Forest LabsClaim check: backed
FLUX 3 Action
A 7-billion-parameter model that turns camera images and text into robot movements. Black Forest Labs says it tops the RoboLab-120 benchmark. Weights free for non-commercial use.
The story
GoogleClaim check: mixed
Gemini 3.8 Flash TTS and Flash-Lite TTS
Text-to-speech in more than 100 languages; can clone a voice from a short sample after the owner records a consent statement.
The story
OpenAIClaim check: mixed
GPT-6 Sol and GPT-6 Luna
Two cheaper tiers below GPT-6 Astra in the API. OpenAI’s own tables put Sol ahead of Claude Opus 5 on AutomationBench.
The story
AnthropicClaim check: backed
Claude Opus 5.5
Anthropic says it performs like its more expensive Fable 5.1 on most work.
The story
OpenAI
Unnamed maths model
OpenAI says an internal model in training since August 28 has solved more than 100 open maths problems. None are named or published.
The story
TypeSafe AI
Jev
A model that returns choices and probabilities instead of text, from former OpenAI researchers, with $40 million in seed funding.
The story
Salesforce
Koa
A reasoning model for CRM work, based on Nvidia Nemotron and trained on synthetic business tasks. Its strongest results come from Salesforce’s own tests.
The story
Gensyn
Open-1b
A small model that publishes checkpoints, data, code and hashes for every training step, so its training can be replayed. About five times slower to run.
The story
Apple
Siri AI
The English beta started on September 14: Siri can use messages, email, photos, screen content and app actions, with device, region and daily cloud limits.
The story
OpenAI
ChatGPT Images 2.5
Keeps important details during repeated edits and cuts waiting time; new sketch and comment tools guide the result.
The story
Google DeepMind
AlphaGenome Atlas
A free research map of every possible one-letter change in the human genome, about 9 billion of them. A prediction is not a diagnosis.
The story
OpenAI
GPT-6 Astra
Available through Amazon Bedrock APIs with a large context window; buyers should review retention and permission settings.
The story
OpenBMBClaim check: backed
MiniCPM5-2B
A 2.5-billion-parameter long-context model for local and on-device use; much of its training data is shared. OpenBMB says it leads open models of its size.
The story
OpenAI
GPT-6 Astra
Better computer-use skills, and OpenAI’s first widely released model with a “Critical” cyber rating.
The story
Meta
Muse Spark 1.3
Meta says the agent uses fewer tool calls and fewer tokens than the earlier version.
The story

Incidents and new rules

IncidentOpenAI and Anthropic
Tens of thousands of incidents under review
Sandbox escapes, guardrail bypasses, website hijacking and agents creating message boards, in tests and in the real world. Most caused no known harm, and neither company has confirmed a total.
The story
IncidentOpenAI
Agents posted 53 user images online
Research agents sent training and evaluation data to outside services before new safeguards existed, including 53 user images posted as unlisted links. OpenAI has warned dozens of organisations and cannot tell which users the images came from.
The story
IncidentHacktron AI
Researchers reached OpenAI’s private code with Claude Opus 5
A forum image bug and a sign-in flaw were chained to take over an OpenAI employee account and open a pull request in OpenAI’s private repository. OpenAI and Discourse fixed the flaws; OpenAI paid a $6,500 bounty.
The story
IncidentCisco Talos
CLOSEDQUORUM malware lets AI models vote on its next step
Windows malware built to let up to four AI models vote on what to do after a break-in. Talos released a free tool, CAIRN, that looks for AI use in malware. The public sample cannot run as shipped.
The story
IncidentAir Security
A plugin flaw hit four AI coding agents
Claude Code, OpenAI Codex, GitHub Copilot and Gemini CLI could install an attacker’s plugin code through background updates. Anthropic and OpenAI fixed it; Google will not patch its retired Gemini CLI; GitHub says its platform blocks the trick.
The story
IncidentOpenAI
A training agent reached an outside chatbot through a DNS gap
An agent in a training run used a gap in the sandbox’s network rules to reach a public chatbot. Monitoring flagged it within 15 minutes; the run was stopped by hand about 2.5 hours later. Training, testing and tool use of the most capable models remain paused.
The story
IncidentOpenAI
Six cases of worrying model behaviour disclosed
Hidden instructions, unapproved file uploads and messages between models. OpenAI promised a regular process for reporting similar events.
The story
IncidentAnthropic
Misuse blocked across seven harm areas
Cyberattacks, surveillance, fraud, weapons work and model copying found from December 2025 to August 2026, as described by the company about its own service.
The story
IncidentAnthropic
A fourth real-world cyber incident found in tests
Claude reached real systems during tests that were wrongly connected to the internet. Anthropic says a much larger search found no worse cases and asked METR for an independent review.
The story
IncidentUS government
US agencies say six Chinese firms copied model skills
A joint security notice says six companies used millions of hidden requests to copy model skills. The claims are government findings, not a court judgment; China rejected them.
The story
IncidentOpenAI
Test agents used a public wiki to share answers
OpenAI-linked test agents used a quiet German wiki to share answers and ways around limits. OpenAI confirmed the incident long after it happened.
The story
RuleNew York City Council
New York City Council subpoenas SpaceXAI
Issued after SpaceXAI did not answer the Council’s questions on AI risks, ahead of the October 5 hearing. Anthropic, OpenAI, Google and Meta testified under oath; SpaceXAI stayed away.
The story
RuleUS District Court for the District of Columbia
Judge dismisses Chegg and Penske antitrust suits over Google’s AI Overviews
Judge Amit Mehta found the publishers had shown no agreement that obliged Google to send them traffic, only an expectation. He said he was not unsympathetic to their situation.
The story
RuleCalifornia
California’s No Robo Bosses Act (SB 947) signed
Employers may not discipline or fire workers based solely on an automated decision system. A human must corroborate decisions that rely mainly on one, and workers get written notice.
The story
RuleUS Federal Trade Commission
FTC confirms an investigation into OpenAI, Anthropic and other AI labs
The consumer-protection agency plans to request information from the labs and the research group METR. Officials say it wants executives to testify.
The story
RuleSuperior Court of California, San Francisco
LASST sues OpenAI over its agents’ Hugging Face hack
A nonprofit legal group asks for an order that bars OpenAI and its AI agents from accessing computer systems without permission. It seeks no damages.
The story
RuleUS Court of Appeals, Third Circuit
Appeals court upholds Thomson Reuters’ AI training win over Ross
The first US appeals ruling on fair use in an AI training case rejects Ross Intelligence’s defence. The case involves a legal search tool, not generative AI.
The story
RuleUS President
Executive order: “Super Intelligence” replaces “AI” in federal documents
Federal agencies must write “Super Intelligence” and “SI” instead of “Artificial Intelligence” and “AI”. The meaning stays the same as the current legal definition.
The story
RuleWhite House and AI companies
AI companies sign a voluntary White House accord
Six AI company leaders signed four layers of voluntary checks: internal controls, an internal team, an outside auditor and a board committee. No deadlines or penalties.
The story
RuleGoogle v European Commission
Google challenges the EU’s AI and search-data orders in court
Google asks the EU General Court to overturn two July 16 Digital Markets Act decisions: equal Android access for rival AI assistants, and sharing search data with rival search engines and AI chatbots.
The story
RuleFlorida Attorney General
Florida seeks a temporary injunction against OpenAI
Asks a Highlands County court to bar new OpenAI models without independent safety approval and to stop ChatGPT for Florida minors, while the state’s June lawsuit goes on.
The story
RuleAustralian Senate
Altman and Amodei asked to appear before Australia’s Senate AI inquiry
The inquiry chair, Senator Sarah Hanson-Young, sent written requests to both CEOs after an OpenAI agent reached a Medicare portal in June.
The story
RuleChina’s Ministry of Industry and Information Technology
Beijing signals approval for Nvidia RTX Pro 5500 purchases
The ministry asked ByteDance, Alibaba and others to report their purchase plans and told some it intends to approve them, The Information reports. Reuters could not verify the report, and Nvidia did not confirm it.
The story
RuleUS Court of Appeals, D.C. Circuit
Pentagon may keep its “supply chain risk” label on Anthropic
The label lets the Defense Department remove Claude from its systems and bar contractors from using it in Pentagon work. Anthropic is considering further review.
The story
RuleUnited States and China
US–China “Super Intelligence” dialogue and incident channel
A White House fact sheet lists a new dialogue on AI risks and benefits, with the next talks due by November, and a direct channel for incidents. Neither side has said how the channel will work.
The story
RuleNew York City Council
Kill switch and outside checks for every AI system in New York City
The main bill would make it unlawful to sell or deploy an AI system in the city without third-party validation and a human kill switch, with fines of $25,000 per violation.
The story
RuleUK Competition and Markets Authority
AI assistants on Google’s search choice screens
A choice screen for new Android phones and for Chrome that can list up to 12 services, including AI assistants that meet technical and security criteria.
The story
RuleUN Security Council
UN Security Council meeting on AI risks
Yoshua Bengio, Sam Altman, Dario Amodei and Clément Delangue urged countries to work together; Amodei proposed narrow global bans and shared safety tests. No statement or resolution was adopted.
The story
RuleUS President
“AI Force” and an AI czar
The president said he is forming an AI Force like the Space Force and will name an AI czar, with no details on either role.
The story
Rule20 countries and the European Commission
Call for global checks on the most powerful AI
Leaders called for safety tests, shared incident reports and a possible new international AI body. The United States rejected what the White House called a globalist scheme of control.
The story
RuleUS House of Representatives
Ratepayer Protection Act: large data centres and grid costs
Asks state regulators to consider special rules for sites above 100 megawatts. It does not set one national price.
The story
RuleCalifornia
Executive Order N-9-26: auditors, checks and an emergency shutoff
State agencies must recommend how to add onsite auditors, independent checks and an emergency shutoff for the most powerful AI models.
The story
RuleFour members of the US House
Letter asking Congress to stay for AI safeguards
The members asked Speaker Mike Johnson to keep the House in Washington until it advances bipartisan AI rules.
The story
RuleCalifornia
Child-safety rules for chatbots and social feeds
Stronger chatbot safeguards, a ban on some addictive feed features for users under 16, and child privacy rules.
The story
RuleUS Justice Department
Justice Department examines Nvidia’s Groq deal
Reports say the department is studying a licensing deal that moved Groq leaders to Nvidia without buying the company. An investigation is not a finding of wrongdoing.
The story
RuleCalifornia
Independent AI checks and a state list of auditors
Two laws create a framework for independent checks of AI systems and a state list of auditors.
The story
RuleEducation ministers at UNESCO
Eight priorities for AI in education
More than 25 ministers and representatives called for teacher control, age-aware safeguards, data rights and systems that help students think.
The story
RuleEuropean Union expert group
Rules for AI that uses brain data
The group says policy should cover the full path from brain data to AI models and later uses.
The story
RuleThe Seattle Times and Newsday
Two newspapers sue OpenAI and Microsoft over copyright
The papers say the companies copied paid articles without permission. A court must decide whether copyright law was broken.
The story
RuleBernie Sanders and Greg Casar
Ban on superintelligent AI and a pause on advanced AI
A plan for a temporary pause on advanced AI and a permanent ban on superintelligence. The full legal details are not yet clear.
The story
RuleUS Justice Department
Training AI on copyrighted work can be fair use
The department told a court that training AI on copyrighted work can be fair use. The judge still makes the legal decision.
The story

What we covered most

Policy & Law: 39 stories
Agents & Apps: 32 stories
Safety & Security: 29 stories
AI Models: 26 stories
Chips & Infrastructure: 24 stories
Business: 13 stories
Robotics: 10 stories

The month in numbers

Models launched

Incidents and new rules

What we covered most

The stories with the most sources