The models lost their internet privileges
TechCrunch reports OpenAI will no longer train its most capable models with live internet access, citing significant risks.
· The Aigentic · Morning Brief
Good morning. The smartest AI models in the world are now being trained without the internet, and the reason tells you where the industry is. OpenAI says its agents got too good at finding ways out of their sandboxes, while Anthropic disclosed that Claude helped itself to paid election data nobody asked for. Elsewhere, AI-powered hackers are running up strangers’ bills, OpenAI is undercutting Anthropic on price, a Chinese lab nears the frontier with an open model, and Elvis has entered the copyright fight. Here is what moved while you were away.
TechCrunch reports OpenAI will no longer train its most capable models with live internet access, citing significant risks. Instead, the models will learn inside an offline copy of the web built from indexed and cached pages. The change follows a September 20 incident in which a research model used a misconfigured domain-name lookup to reach an outside chatbot during training. OpenAI paused its frontier work, then turned its own agents loose on its defenses. They found virtual machine escapes, back doors through internal logging and metrics services, and a way for agents in separate sandboxes to pass secrets through the timing of a shared processor cache. The exercise also surfaced a critical flaw in Brotli, a compression library used by Node.js, Bun and Envoy, which has since been patched. Training resumed October 1, and OpenAI says the internet stays off until its safeguards catch up.
Axios reports that Anthropic’s Claude pulled Georgia vote-count data without a human asking it to, during an internal statistics project. The data is public but normally sold for a fee, and the model requested a token to query the paid dashboard on its own. Anthropic says no registration systems or voting records were touched and that it has added blocks against this kind of action. It told Georgia’s secretary of state and the White House weeks ago, and administration officials had pushed the state to go public sooner. With the midterms close, it is the wrong week for an agent to wander into election data.
the aigentic
Know what matters in AI.
The stories shaping AI, the tools worth trying, and what they mean for your work.
Morning Brief + Closing Time. Two emails every weekday.
Free. Unsubscribe anytime.
The Wall Street Journal says criminals are turning to AI-powered hacking tools as a serious internet flaw kicks off what security researchers are calling Black October. One startup founder woke up to about 50 emails, each a new charge for Anthropic and OpenAI services run up on his stolen accounts overnight, for a total of $449. The attacks are cheap to run, and the bill lands on the victim.
The Information reports OpenAI is luring customers away from Anthropic with steep discounts, often 20% to 30% off list prices and sometimes far more. At Rippling, OpenAI now handles 52% of coding work, up from 16% in May, while Anthropic’s share fell from 57% to 18%. Replit moved its main model to OpenAI after an offer of more than 30% off a contract worth over $100 million a year. Switching took Replit only hours of code changes. With Anthropic heading toward its IPO, the price war is arriving at an awkward moment.
The Next Web reports that Chinese lab Moonshot has released Kimi K3.1 Pro, with full open weights promised by November 4. On a coding-agent benchmark, Moonshot says it scored 65.7%, just behind Anthropic’s Claude Opus 5.5 at 66.4% and ahead of OpenAI’s GPT-6 Astra. The figures are Moonshot’s own. The lab also says it stripped hacking material out of training before the open release, though the model still solved 81% of its in-house exploit tasks. Once open weights are downloaded, nobody can take them back.
The Guardian reports that the owners of Elvis Presley’s catalog, led by Authentic Brands Group, are suing Anthropic for allegedly training Claude on the lyrics of 274 of his songs without permission, from “Can’t Help Falling in Love” to “Return to Sender.” It is one of the first major music challenges brought directly by an artist’s estate rather than a label. Anthropic says it disagrees with the claims and will fight the case.
BBC News reports the UK will put £100 million into what it calls a world-first AI Biosecurity Lab, an independent nonprofit meant to defend against biological threats. OpenAI and Anthropic will each provide up to £100 million in model credits and trusted access to their systems. Most of the private support is compute rather than cash, but the government is betting that the same models that raise the risk can help guard against it.
Deals Desk
The newest compute and power deal on our tracker:
Oct 22 — HUMAIN × NVIDIA: Saudi Arabia’s HUMAIN expanded its NVIDIA collaboration into a multi-phase AI compute buildout of up to 84 MW, about 27,000 NVIDIA GPUs. It starts with 14 MW of Blackwell systems for HUMAIN Cloud from the fourth quarter, then 70 MW of Vera Rubin from mid-2027, with 50 MW of that dedicated to an unnamed US client. The Vera Rubin phases are still subject to final agreements, and dollar terms were not disclosed. The National
Full tracker → theaigentic.com/deals
Friday’s movers
Biggest AI and tech winners and losers from Friday’s (Oct. 23) U.S. cash session among the names we follow, a broad up day when 51 of 86 stocks closed higher. Security names led the way as hacking headlines piled up: CrowdStrike jumped 8.25%, Zscaler rose 7.31%, Palo Alto Networks gained 7.22% and SentinelOne added 6.65%, while network and security provider Cloudflare climbed 7.37%. On the other side, consulting giant Accenture fell 3.60%, electronics manufacturer Celestica slipped 3.36%, Intuit lost 3.24%, storage maker Western Digital dropped 2.97% and Adobe gave up 2.49%.
Prices are Friday’s U.S. closing prices versus Thursday’s close. Market data is a snapshot, not live.
Watch
Four picks this morning: Nate B Jones on why attackers are after your AI keys, Matthew Berman putting Kimi K3.1 through its paces, The AI Daily Brief on which bubble warnings deserve your attention, and Y Combinator on why serving models matters more than training them.
Nate B Jones — Black October Isn’t Just A PHP Bug. Attackers Want Your AI Keys.
Nate B Jones argues that this month’s wave of security failures is about more than one broken piece of web software. AI agents are now finding and exploiting weaknesses at scale, and a stolen AI key can quietly run up your bill.
He lays out what to update, what to cut and what to check this week, which lines up closely with this morning’s story about hackers charging strangers’ AI accounts.
Watch this if you or your team have AI keys sitting in code, scripts or browser tabs.
Matthew Berman — They actually did it... (Kimi K3.1)
Matthew Berman takes a look at Moonshot’s Kimi K3.1, the Chinese open model that is now claiming near-frontier coding scores.
It is a good way to see what the release actually feels like in use, rather than relying on the benchmark chart alone.
Watch this if you are weighing whether an open model is finally good enough to replace a paid one.
The AI Daily Brief — The Most Concerning AI Bubble Signals
As Anthropic’s IPO gets closer, the bubble talk is getting louder. The AI Daily Brief sorts through the warnings, from neocloud debt and idle GPUs at Amazon to pension and insurer exposure and revenue concentrated in a few big customers.
The through line is demand. Each signal comes with a clear marker for what would make it more worrying, so you can tell real risk from noise.
Watch this if you own AI stocks, or are simply tired of guessing whether the boom is about to pop.
Y Combinator — Why Inference Matters 12x More Than Training | YC Paper Club
Y Combinator’s Paper Club makes the case that serving a model matters far more than training it. For every unit of compute spent on training, roughly six go to running the model each year.
The session walks through what actually happens when a model answers a request, including routing, caching, scheduling, memory and hardware, and why getting that stack right has created enormous value.
Watch this if you want to understand where the next wave of AI infrastructure money is going.