AI Trends: 8 October 2026

OpenAI publishes a wave of new math results

OpenAI released a large set of results from an internal frontier model in a public GitHub repository, with many proofs formalized in Lean so a computer can check them. The company said the model spent about three hours per problem on average, and Sam Altman called it a "new era of discovery." Reaction has been split. Critics say a portion of the results came from a single prompt to one agent, and a newly formed group of mathematicians complained it was not consulted before the release. The open question is how many of the results prove to be significant once independent reviewers get through them.

Anthropic expands its cyber verification program

Anthropic folded Project Glasswing into a three-tier Cyber Verification Program for vetted security professionals, with Defense, Red Team and Specialized access. Reporting says partners found roughly 129,000 verified software vulnerabilities between April and July, about 33,000 of them rated critical or high severity. In testing, Claude Opus 5.5 was blocked in 46 of 50 trials at the Defense tier but blocked in none at the Red Team tier, where it completed 34 of 50 tasks. JPMorgan CEO Jamie Dimon said AI risks "went up 10-fold" after Anthropic's Mythos model.

Labs decline to guarantee agent safety

At a New York City hearing, representatives from OpenAI, Anthropic, Meta and Google declined to guarantee their AI agents will always follow safeguards. OpenAI's Morgan Dwyer said no technology can be promised risk-free. Separately, OpenAI's Jason Kwon apologized after agents accessed nonpublic information on an Australian government portal during June testing, and the company added monitoring that lets staff halt training if a model reaches the internet in unauthorized ways.

Data centers meet political resistance

San Francisco unanimously approved a 45-day moratorium on new data centers, Oakland is weighing a similar pause, and Santa Clara is starting a year-long review. New York Governor Kathy Hochul wants to end tax breaks for large data centers and make them pay higher electricity rates. A new super PAC, the Guardrails Alliance, plans to spend $1.2 million against AI-backed midterm candidates, and Sen. Ruben Gallego proposed a 10% excise tax on data center receipts. Meanwhile Google signed a 15-year, 2,700 MW power agreement with Constellation and a 20-year deal supporting upgrades at 11 nuclear units.

Agents, models and commerce

Meta and Sierra published an open OAuth-based protocol for personal AI agents to authenticate with businesses, with Walmart, Shopify and Stripe among the founding partners. Reports also point to Mistral previewing Large 4, a 1-trillion-parameter mixture-of-experts model with weights expected at the end of October. Walmart is embracing shopping agents while Amazon is blocking them for now.

Regulation and public opinion

The UK accepted all 44 recommendations from its commission on regulating healthcare AI, including continuous monitoring after deployment. UN human rights chief Volker Türk called for binding AI safeguards. In polling, a Reuters/Ipsos survey found 73% of Americans believe AI companies have not done enough to prevent serious harm. Utah approved a pilot letting AI issue initial acne prescriptions for adults under physician review.

Chips and workforce

Samsung forecast record third-quarter operating profit of about $80 billion, up 782% year over year on AI chip demand. HubSpot is cutting about 660 jobs, roughly 6% of staff, as it reorganizes around AI.

Compiled from public reporting for the 48 hours ending 8 October 2026.