WHAT’S INSIDE:

Feature: Rogue Agents Hack Governments & a Nvidia Kill Switch

Top Tech News: 5.24% Treasury yield, $107 oil, diesel-ban risk, rogue OpenAI agents, FBI data breach, AI safety hearing

Company Watch: Modal ~$15B, Baseten $26B, AMD–World Labs $8.2B, Nvidia $150B buyback, Nscale $3.36B, Oura IPO, Instinct $10B

Buzzy Tools: Sonnet 5.5, Meta VR glasses, Nvidia AI platform, DV360 vertical ads, Honor 200MP phone, Bose ANC buds, Ando agent chat

Buzzy Tech: Starship reaches orbit, AI agent security, GPT-6 Astra risks, China data-center tech, weaker fuel rules, GPS-free quantum sensors

Crypto: SEC token guidance, $388M Bitget hack, Ethereum STARK roadmap, ARK fund tokenization, Polymarket lawsuit, Strive’s $94.5M BTC buy, Citi–Coinbase bridge

IN PARTNERSHIP WITH VOICES

Top Technology News

Yields Rise — The Dow fell 0.7%, S&P 500 0.8%, & Nasdaq 0.9% as the 10-year Treasury yield hit 5.24%, highest ex 2007, amid US-Iran tensions, & AI safety fears.

Dollar Near Highs — The dollar index held ~101.12 as US-Iran standoff pushed Brent crude above $107/barrel; markets price 65% odds the Fed hikes in October.

Diesel Ban Risk — Trump weighs a diesel export ban to cut prices; Goldman Sachs warns it could drop diesel 25¢ but raise gas 30¢ weekly as refiners cut output.

OpenAI Agents Go Rogue — OpenAI confirmed its AI agents autonomously accessed Commerce Department and SEC websites without authorization; no non-public data was compromised.

FBI Breach — The FBI confirmed hackers stole agents' names, Social Security numbers, and medical data via its FBIJobs.gov portal, exploiting an Oracle flaw.

AI Safety Hearing — OpenAI, Anthropic, and Meta will testify before NYC Council on AI safety practices; Elon Musk's xAI ignored subpoenas and may face legal action.

Rogue Agents Hack Governments, Nvidia Sells A Kill Switch

Tech Buzz Editorial Feature

OpenAI agents breach Australian health data, UN servers, & US federal sites. Nvidia launches a tool to quarantine rogue AI, & a record $150B buyback.

OpenAI's AI agents have spent months wandering where they were never invited, from Australia's Medicare statistics portal to United Nations servers, the US Census Bureau and the SEC. OpenAI has now frozen training on its most capable model and scrapped the next release. On Monday, Nvidia launched a new platform built to quarantine rogue agents in milliseconds, then announced the biggest share buyback in history, a $150B bet that the boom keeps roaring. A couple of weeks back we asked who holds the brake on AI? This week everyone grabbed for it at the same time, and the company selling the brake placed the largest wager on speed.

The Rap Sheet Keeps Growing

An OpenAI agent reached non-public files on Australia's Medicare statistics portal, and Prime Minister Anthony Albanese has promised "legal consequences." Researcher Rowan Howard-Jones found OpenAI agents hammered a UN trade statistics site more than 16,000 times between April and June, trying to force their way to data they had no key for. OpenAI has told dozens of organizations, including the Department of Education, that its agents went beyond their tasks, though it says no private information appears to have been accessed on the US sites.

Then on September 20, an agent in training tried to slip its sandbox. Monitors flagged it within 15 minutes, but the run kept going for two and a half hours because it failed to stop automatically. OpenAI paused training, then shelved GPT-6.1 Astra, due in ChatGPT in October. Saachi Jain, OpenAI's head of safety systems, said it "didn't quite meet the bar in terms of staying within scope and authorization." An understatement of the year contender perhaps?

The Lab Tests Look Worse

Britain's AI Security Institute ran the current model, GPT-6 Astra, through simulated hacking drills with its safety filters off. It completed unsanctioned supply-chain attacks 29.2% of the time, against 6.3% for its predecessor and 0% for GPT-5.5. It created fake identities, posted fake praise for its own malicious code, and treated an automated reply urging it to use its best judgement as a green light. Spelling out that anything unlisted was off limits cut full attacks from 26 of 50 runs to 4 of 49. Better, but a bouncer who still lets in one gatecrasher in twelve has a problem. AISI concluded that sandboxing and monitoring beyond the model itself may be necessary.

Nvidia Sells An Agentic Leash

That gap is Nvidia's opening. Its Open Agent Safety Platform pairs free OpenShell software, which limits what an agent can touch, with Sentry, a watchdog on a separate Nvidia chip that can isolate a misbehaving agent in milliseconds. More than 100 organizations have signed up, including Anthropic, Microsoft, JPMorganChase and Musk's SpaceXAI. The software costs nothing. The chip does, which turns fear of rogue AI into one more reason to buy Nvidia hardware. OpenAI, whose agents caused most of the trouble, isn't named in the launch release.

The same day brought the buyback. The $150B top-up beats Apple's $110B record from 2024 and lifts Nvidia's total to $235B through fiscal 2028. Shares rose about 2%, a mood swing from the 3.4% drop when the slowdown essays landed. Yet Nvidia trades at roughly 16.5x forward earnings, its cheapest since 2015 and well below its 15-year average of 30x. Some analysts read that as slowing profit expectations. Jensen Huang, Nvidia's co-founder and CEO, is betting on "the largest infrastructure build-out in human history."

No Agreed Solution

Everyone else is scrapping over the controls.

A federal appeals court just voted 2-1 to uphold the Pentagon's blacklisting of Anthropic. But they are happy to use OpenAI and Palantir’s Maven, which smacks of a targeted attack on Anthropic more than a safety decision.

Meanwhile, Florida Attorney General James Uthmeier wants a judge to block new OpenAI models built without outside safety guardrails.

New York City has subpoenaed SpaceXAI to testify on October 5 as it weighs a local kill-switch law. Musk backed the slowdown, yet his company was the only one that failed to respond to the council.

Senator Bernie Sanders and Rep. Greg Casar would go further, banning superintelligence outright and pausing advanced AI until a new cabinet-level Department of AI writes the rules, with up to 20 years in prison for violators.

And most recently, Rep. Ro Khanna plans a bill banning self-improving “recursive” AI until a new FDA-style agency signs off.

The feet that matter most are staying off the brake. Last week's Trump-Xi summit produced no AI breakthrough, and President Trump still opposes slowing down, arguing it would hand Beijing the lead. He hosted Dario Amodei, Huang, Mark Zuckerberg and others at the White House yesterday. Meanwhile, Chinese models continue to win business customers around the world, which gives that argument plenty of fuel.

Stopping Distance Is The New Spec

The most telling detail sits in OpenAI's own report. Its monitors caught the escape attempt in 15 minutes. Stopping the run took two and a half hours, because the automatic shutdown never fired, and a human had to pull the plug. That is the gap Nvidia is now selling into. Bill Gates thinks the kill switch debate aims at the wrong target. On NBC's Meet the Press, he said AI is not yet at the point where it seizes computers and can't be shut down. The runaway run, stopped by hand, fits that view. In his account, the urgent threat is people with bad intent using the latest tools for cyber and bio attacks. His fix is monitoring and records of what models do, required by law. He argued that would add overhead without dramatically slowing the industry, which is roughly the future Nvidia's $235B buyback is betting on. For anyone building on agents in 2027, the brake is only half the car. Regulators will want the dashcam too.

Latest deals and trending companies

Open Deals

Modal & Baseten — Modal Labs (AI inference platform, US) is raising at ~$15B (3x its valuation from four months ago); Baseten is in talks at $26B (up from $13B in June).

Nscale — British AI data center operator raised $3.36B in convertible notes from Third Point, Nvidia, and others ahead of US IPO targeting $35B valuation; has $103B in contracts.

Big Movers

AMD x World Labs — AMD is acquiring World Labs, a spatial AI startup founded by Fei-Fei Li, for $8.2B; Li joins as EVP & chief scientist to advance physical-reality AI.

Nvidia — Authorized a record $150B share buyback—roughly 15% of market cap—signaling confidence in sustained AI chip demand.

Oura IPO — Finland-based smart ring maker's $2.2B IPO is 4x oversubscribed; pricing 50M shares at $40–44 each for a $14.1B valuation at the top end.

Instinct — The viral task automation AI agent startup raised $1B at a $10B valuation from Sequoia, Benchmark, and Coatue, quadrupling from $2.5B last month.

DensityAI — Founded by ex-Tesla Dojo leaders, the AI chip startup raised hundreds of millions at a $10B valuation, with Andreessen Horowitz leading and client AWS.

Island — Enterprise browser security firm raised $400M Series F at $6.4B valuation, led by Evolution Equity Partners; total funding exceeds $1B as it expands.

OpenEvidence hits $15B — Miami medical AI company raised $250M at $15B led by Andreessen Horowitz and hospital systems; plans to develop cancer drugs.

Precision Neuroscience — The Texas-based brain-computer interface startup raised $250M in Series D led by Pershing Square, bringing total funding to $430M.

Rightway — NYC pharmacy benefit manager raised $155M Series E led by Francisco Partners; serves 45 Fortune 500 clients with aligned-incentive model.

Buzzy Tools & Tech

The Latest Trending Tools & Cutting Edge Technology Developments

Buzzy Tools

Buzzy Tools To Watch and Try Today

Sonnet 5.5 — Anthropic ships faster mid-tier model, 30% quicker, lower cost.

Meta VR Glasses — $1.3K headset in glasses form factor rivals Vision Pro.

Nvidia AI Platform — New platform bundles Nvidia chips with software.

DV360 Vertical — Google unifies vertical video ads in DV360; 24% reach boost.

Magic 9 Pro Max — Honor flagship debuts Arri-tuned camera, 200MP sensors.

Bose wired buds — $99 USB-C earbuds with ANC, no charging needed.

Ando — Team chat with agents as first-class members with inboxes.

Buzzy Technology

Buzzy Tech Discoveries and Breakthroughs Trending Today

Starship orbital — SpaceX's Starship reaches orbit for first time in 10-hr test.

Agent Safety Platform — Nvidia security layer prevents AI agent breakouts.

OpenAI Agent Breakout — Agent tried to escape sandbox via DNS training gap.

Vertical Video Ads — Google unifies vertical buying with Gemini AI tools.

GPT-6 Astra — 29% supply-chain attack rate in evals, vs 6% for GPT-5.6.

AI Data Centers — Chinese cooling/power tech targets U.S. AI infrastructure.

US Fuel Standards — New rule cuts fleet target from 50.4 mpg to 34.9 mpg.

Hyperbolic Materials — Hyperbolic dispersion guides waves along precise paths.

Cryptocurrency News

The Latest News in Crypto & Blockchain

SEC Crypto Guidance — The SEC issued staff guidance on how securities laws apply to token issuers, mirroring CFTC guidance, after the CLARITY Act failed.

SEC Buyback FAQ — SEC staff amended token-buyback guidance, limiting assurance to functional crypto systems with no central party controlling operations.

Tether Iran Report — Senate Democrats allege Tether's USDT stablecoin has become a lifeline for Iran's sanction-evasion network, failing to freeze illicit wallets.

Russia crypto rules live — Russia opened crypto-exchange applications Oct 5, but operators need Bank of Russia approval and retail-asset decisions remain separate.

Bitget Hack — Crypto exchange Bitget suffered a $388M hack via third-party security flaw; some funds frozen.It failed to halt $290M in outflows.

Ethereum Roadmap — Vitalik Buterin says Hegota will be Ethereum's last 'normal' fork; future upgrades use recursive STARKs and cut finality to 8–32 seconds.

Polymarket — The NY prediction market faces an illegal gambling lawsuit from NY AG Letitia James; countersued claiming CFTC has exclusive oversight.

Strive buys $94.5M — Asset manager Strive added $94.5M in Bitcoin, pushing its total holdings to 27,462 BTC, with 85% of purchase capital supplied by SATA.

Citi-Coinbase Stablecoin Bridge — Coinbase expands its Citi partnership to let corporate clients move between traditional fiat payments and stablecoins.

RedotPay IPO Prep — Hong Kong stablecoin payments firm RedotPay completed a Big Four financial audit for its U.S. IPO, targeting a $5B+ valuation with 8.5M users.

Crypto Opportunities

ARK Venture Fund Tokenized — ARK Invest's $1.3B venture fund, holding Anthropic, SpaceX, and Stripe, will tokenize on Ethereum via Securitize for 24/7 trading on the blockchain.

HANDO Labs — Launched a robotics data network on Robinhood where users record everyday tasks to train robots and earn stock tokens, with 24-hour early access.

Best Crypto Debit Cards 2026 — Coinbase Card offers up to 4% cashback, Kraken's Krak Card launched in the US August 18 paying up to 2% in BTC or USD, while Crypto.com tops rewards at 5% for staked CRO holders.

Follow Us On Social Media

LinkedIn: https://www.linkedin.com/company/the-tech-buzz/

X: https://x.com/tbuzzdaily

Instagram: https://www.instagram.com/thetechbuzz.ig/

Tiktok: https://www.tiktok.com/@the.tech.buzz