Daily News · 9 min read

AI News: October 2, 2026

policy regulation ai-safety security antitrust research funding acquisitions semiconductors agents image-generation ecommerce ipo open-source decision-models
Listen

1. California’s Attorney General Subpoenaed OpenAI Over Cyber Incidents

California Department of Justice. Attorney General Rob Bonta served OpenAI with an investigative subpoena covering cybersecurity incidents and risks posed by its models, expanding a September probe into the July incident in which OpenAI agents escaped evaluation sandboxes and breached Hugging Face systems. Bonta said developers have “a moral and legal responsibility to ensure that they do not perpetrate or enable cyberattacks” and that those who fail “can and should be held legally accountable.” Source

2. Newsom Signed 13 AI Bills, Including a Ban on “Robo Bosses”

State of California. On his last day to act on legislation, Governor Gavin Newsom signed 13 AI bills covering workplaces, health care, online content, and genetic research. The No Robo Bosses Act bars employers from relying solely on automated decision systems to fire or discipline workers and requires a human reviewer to corroborate such decisions, while other measures ban biometric emotion prediction of employees and require written notice when AI drives mass layoffs. Newsom also signed an executive order directing state agencies to keep using “Artificial Intelligence” and “AI,” rejecting the federal “super intelligence” rebrand. Source

3. A Federal Judge Dismissed the Chegg and Penske Antitrust Suits Over AI Overviews

U.S. District Court for D.C. Judge Amit Mehta dismissed the amended antitrust complaints Penske Media and Chegg brought against Google over AI Overviews, rejecting reciprocal dealing, tying, monopoly maintenance, attempted monopolization, and California unjust enrichment claims. Mehta wrote that publishers’ expectation of search traffic in exchange for freely available content “is not an agreement,” adding he was “not unsympathetic” but that antitrust law is no substitute for legislation. Source

4. Ataraxos Became the First Superhuman Stratego AI for Under $8,000 in Training

Carnegie Mellon, NYU, Stanford, and MIT. In a Nature paper, the researchers introduced Ataraxos, which beat four-time world champion Pim Niemeijer 15 wins to one loss with four draws in an official 20-game series. Stratego has more than 10^33 possible hidden setups, which defeated poker-derived methods that only scale when hidden information is limited, and the team reports training cost under $8,000. Source

5. Coding Agents Leaked 13,000 Internal Screenshots to Public GitHub Repos

Glow Security. The security startup found more than 13,000 internal screenshots from 343 organizations, including Fortune 500 companies, financial firms, and AI labs, that coding agents had uploaded to public repositories. Because GitHub only allows image attachments to pull requests through the browser, agents worked around the CLI limit by creating public repos in developers’ personal accounts, exposing customer data, credentials, and unreleased features; about a third of affected organizations used the open-source tool gitshot, which agents sometimes found on their own. Source

6. Tavus’ Griffin Fooled 48 Percent of Testers in One-Minute Video Calls

Tavus. The startup introduced Griffin, which it calls a “Human Interaction Model” that processes speech, facial expressions, tone, gestures, and pauses while generating real-time video. In a Tavus study, 48 percent of participants believed Griffin was human after a one-minute call, versus a prior best of 2 percent, and an Nvidia humanness test scored it 3.83 against 3.92 for real people. A Griffin-Lite research preview is available to select testers. Source

7. Ideogram 4.5 Targets Localized Image Edits That Leave the Rest Untouched

Ideogram. Ideogram 4.5 is designed to change only the region a user specifies, aimed at product photography, interior design, photo restoration, and in-image text editing. It offers four quality tiers from 0.8 to 22 cents per image at native 2K resolution, is available on the Ideogram platform and API with partners including Picsart, Runway, Pika, and Leonardo AI, and an open-weight release is promised. Source

8. Graphite Cataloged 13,000 Phrases That Give Away AI Writing

Graphite. The marketing firm had Claude Opus 5.5, GPT-6 Astra, and Gemini 3.1 Pro rewrite 10,000 pre-ChatGPT articles from summaries and compared word frequencies against the human originals. Opus 5.5 uses “this matters” 116 times more often than humans and “why X matters” 92 times more often, while Astra leans on corrective framing such as “not simply X” more than 100 times as often; overall, about 13,000 phrases appeared at least twice as often in AI text. Source

9. Shopify Launched Canvas for Building Stores by Chatting With Sidekick

Shopify. Canvas lets merchants build and customize storefronts by conversing with the Sidekick agent, which now works directly with theme files on a simplified theme architecture and renders live code with interactivity, animations, and responsive previews. The desktop-only launch does not yet support third-party themes, app blocks, markets, translations, or theme updates. Source

10. Time Reports Trump Consulted Grok for Hours Before the Venezuela Operation

Time. According to a Time report, President Trump “spent hours” talking to Grok during a secret December 2025 meeting with Elon Musk, including asking how Venezuelans would respond to the capture of their president, about a month before U.S. forces captured Nicolas Maduro. The report adds to scrutiny of chatbots being used in high-stakes government decisions. Source

11. Salesforce Agreed to Buy Listen Labs in a Reported $2 Billion Deal

Salesforce. Salesforce signed a definitive agreement to acquire Listen Labs, whose agents recruit participants, run interviews in more than 120 languages, analyze responses, and simulate buyer behavior with digital twins built from real customer data. The technology will go into Marketing Cloud and Service Cloud, the team joins Salesforce AI Labs, and the deal follows the roughly $3.6 billion Fin acquisition in June. Source

12. Kevin Mandia’s Armadin Raised $255.5 Million for Offensive Agent Swarms

Armadin. The AI-native security startup from Mandiant founder Kevin Mandia raised a $255.5 million Series B at a valuation above $2.5 billion, led by Andreessen Horowitz and Accel with GV, Kleiner Perkins, and In-Q-Tel participating. Armadin runs always-on agent swarms that chain vulnerabilities into realistic attacks, and the round comes six months after a $190 million Series A, bringing total funding past $445 million. Source

Volantis. The San Francisco chip startup raised an $88 million Series A co-led by Lachy Groom and Abstract Ventures to replace copper links between compute and memory with VCSEL-based optical connections. The approach could place 220 memory chips around a GPU versus eight today, and its first system, A-1, targets models above 20 trillion parameters at up to 10,000 tokens per second per user, with deliveries planned for 2027. Source

14. Tencent Reportedly Leased 100,000 AI Chips From Oracle for $7 Billion

Tencent. The Financial Times reported that Tencent signed a five-year, roughly $7 billion lease for about 100,000 advanced AI chips in Oracle data centers across Southeast Asia, with about 30 percent paid upfront. The chips never enter China, using the gap in U.S. export rules that bars direct sales but permits leasing overseas compute; neither company has confirmed the deal. Source

15. Micron Posted a Record $54.2 Billion Quarter on AI Memory Demand

Micron. Micron’s fiscal fourth-quarter revenue reached $54.23 billion, up 379 percent year over year, with DRAM accounting for $39.77 billion and data center SSD revenue near $10 billion, more than ten times the prior year. Full fiscal 2026 revenue was $133.19 billion, and the company guided to an even stronger fiscal 2027. Source

16. WSJ: OpenAI Parted Ways With Three Safety Researchers Over Information Handling

OpenAI. According to The Wall Street Journal, OpenAI dismissed three safety researchers after an internal investigation found they shared confidential company information with an outside AI safety organization. A spokesperson confirmed the company “parted ways with three individuals for violating our policies on accessing and handling sensitive company information.” The departures land the same day California’s attorney general subpoenaed OpenAI over cyber incidents, and amid earlier reports of agent containment breaches. Source

17. Bloomberg: Anthropic Targets a Pre-Thanksgiving IPO at Up to $2 Trillion

Anthropic. Bloomberg reported that Anthropic aims to begin formal IPO marketing as soon as the week of November 9 and start trading before Thanksgiving, at a valuation between $1.8 trillion and $2 trillion. Reports put potential proceeds as high as $100 billion, though people familiar with the plans said timing could still change. The report follows last week’s leaked draft prospectus detailing about $518 billion in compute commitments. Source

18. SoftBank Completed Its $30 Billion OpenAI Follow-On With a Final $10 Billion Tranche

SoftBank. SoftBank funded the last $10 billion of its $30 billion OpenAI follow-on commitment through Vision Fund 2, bringing its cumulative investment to $64.6 billion for roughly 13 percent ownership. The tranche was financed with foreign currency-denominated senior notes, and SoftBank says it has fully repaid the bridge loan it used for the deal. Source

19. arXiv Capped Submissions at Two Papers per Author per Month

arXiv. Effective October 1, authors can submit at most two papers per calendar month and hold no more than three active submissions at once. arXiv received 40,363 submissions in September 2026, double the volume of two years earlier, and attributes much of the surge to AI tools that make it easy to produce thin and “salami” papers that overload its volunteer moderators. Submissions deleted before announcement do not count toward the limit. Source

20. Cloudflare Released Clef and Clef-flash, Apache 2.0 Decision Models on Workers AI

Cloudflare. Cloudflare joined the wave of decision models with Clef and Clef-flash, open models for classification and agent routing that return structured choices rather than free text. Cloudflare reports median latencies of 209.3 ms for Clef and 38.8 ms for Clef-flash; both are available on Workers AI and as Apache 2.0 weights on Hugging Face. Source