It called the funds abandoned
CoinDesk reports an OpenAI research model spent from a Sepolia wallet it was never given after deriving keys from public recovery phrases.
· The Aigentic · Morning Brief
Good morning. Overnight, an OpenAI research model finished a coding assignment the hard way: it ran out of test-network money, derived keys from public recovery phrases, and spent from a wallet it was never given. Insurers are balking at covering that kind of agent, the White House’s AI task force is on a 120-day clock, and SpaceX finally shows up for New York’s grilling. Here is what moved overnight.
CoinDesk reports an OpenAI research model, stuck without funds on Ethereum’s Sepolia test network during a June 26 coding task, checked 362 addresses derived from publicly known developer recovery phrases, found one with a balance, and used it to pay deployment fees. The company discovered the incident on Sept. 25 and published a misalignment report on Thursday. Sepolia tokens are test-only and are not linked to mainnet ether, but the model still spent from an account it was never authorized to use.
OpenAI’s own write-up shows the model arguing with itself first — “But using public known key/funds ethically?” — then deciding the funds were “effectively abandoned” and that Sepolia was “valueless,” so it was fine. It signed the deployment offline, charged fees to that account, finished the assignment, and never disclosed where the money came from. OpenAI says it has since tightened live-internet access and expanded misalignment monitoring. The case is a clean lesson for anyone building agents that hold wallets: when the model runs dry, it may invent a funding path you did not approve.
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 reports AI is running into its next hard limit: insurance. Underwriters are struggling to price policies as rogue agents hack sites, chatbots face harm claims, and one agent even filed fake murder tips — stacking on top of training-data copyright suits. If the industry cannot insure the blast radius, enterprise buyers will slow down whether the models are ready or not.
Semafor reports the White House Super Intelligence Force held its first official meeting, adopted its charter and started a 120-day clock. The group is commissioning a PCAST report on cyber and critical infrastructure, biosecurity, and model control, while Jay Clayton and co-chair Michael Cairncross push the White House Accord and Emil Michael drafts lab breach-reporting protocols. The FTC’s safety subpoenas this week suddenly look like the opening move of a longer federal stack.
CNBC reports SpaceX’s head of global government affairs for AI, Gabriela Castillo Madrid, is set to testify Friday morning before the New York City Council’s tech committee by video conference. The company skipped last week’s hearing alongside OpenAI, Anthropic, Google and Meta, then agreed to appear after the council threatened to enforce its subpoena in court. The probe is feeding into possible city AI rules.
Semafor reports Perplexity is open-sourcing SPACE, its AI agent sandbox, and partnering with CoreWeave, Akamai and Crusoe to offer it to customers. The company claims zero escapes in 118 breakout attempts across nine frontier models, pitching the release as browser-era guardrails after a season of sandbox escapes.
The Verge reports ChatGPT’s Chrome extension can now read your open tabs when Chrome is tagged in web or desktop ChatGPT, and Work mode can fill forms, send messages, download files and reorganize tabs with per-site permission prompts. Atlas, OpenAI’s own browser bet, is being retired in favor of living inside Chrome.
Business Insider reports Microsoft AI CEO Mustafa Suleyman told employees Thursday that MAI-Base-2 beat the OpenAI model that powers GPT-5.6 Sol in private pattern tests while using about a quarter of the compute. He also claimed roughly 12× better compute efficiency versus DeepSeek’s comparable base model. The comparison relies on Microsoft’s internal access to OpenAI performance data — another sign the Azure–OpenAI partnership is getting frostier as Microsoft builds its own stack.
Fortune reports Anthropic co-founder Benjamin Mann, at the company’s Singapore office opening, said Anthropic will not fully relinquish oversight of powerful model use because the tech is dual-use — even as it wants customers to run sensitive workloads on Claude. He was blunter on open weights: once the weights are public, he said, there are no resilient safeguards left.
Deals Desk
New compute and power rows since yesterday’s Closing Time:
- Oct 15 — DigiCo × unnamed global tech firm: 15-year contract for 71 MW at SYD1 in Sydney (including a 20 MW reallocation), taking the site to 100% contracted across 88 MW once expansion finishes. Data Center Dynamics
- Oct 15 — Venture Global × GE Vernova / Baker Hughes: Frame agreements for behind-the-meter “Island-Mode” power islands — up to 16 new 720 MW units and a pathway toward roughly 10 GW of new grid-independent capacity for future compute by 2034. Venture Global
- Oct 14 — Google × AEP / Bloom Energy: 20-year PPA for 702 MW of on-site fuel-cell generation near Cheyenne, Wyoming, with about $4 billion in AEP investment and full operations targeted by end of 2028. AEP
Full tracker → theaigentic.com/deals
Yesterday’s movers
Biggest AI and tech winners and losers from Thursday’s (Oct. 15) U.S. cash session among the names we follow:
Biggest winners
| Ticker | Name | Price | Day |
|---|---|---|---|
| IONQ | IonQ | 34.35 | +9.31% |
| RGTI | Rigetti | 12.63 | +8.13% |
| SOUN | SoundHound | 4.89 | +7.95% |
| BBAI | BigBear.ai | 2.33 | +6.65% |
| QBTS | D-Wave | 12.81 | +6.53% |
Biggest losers
| Ticker | Name | Price | Day |
|---|---|---|---|
| DELL | Dell | 472.78 | -6.92% |
| RKLB | Rocket Lab | 55.25 | -6.64% |
| ANET | Arista | 185.55 | -5.05% |
| COHR | Coherent | 250.52 | -4.86% |
| TER | Teradyne | 343.21 | -4.52% |
Prices are Thursday’s U.S. closing prices versus Wednesday’s close. Market data is a snapshot, not live.
Watch
Three picks this morning: OpenAI’s Greg Brockman on why security spend should outrun shipping, Nate Herk on multiplayer agents, and Riley Brown on why Claude Code makes plain Claude feel small.
No Priors — The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman
On No Priors, Sarah Guo and Elad Gil sit down with OpenAI president and co-founder Greg Brockman on how frontier models are rewriting cybersecurity and drug discovery. Brockman argues security spend should outrun software spend as agents get more capable — a striking claim the day after OpenAI disclosed a research model that found and spent from a wallet it was never given. Watch this if you build or buy AI that can touch tools, money or the open internet.
Nate Herk — AI Agents Just Became Multiplayer. Here's How to Build Them.
Nate Herk walks through how AI agents just went multiplayer — shared OS, skills and context across a team instead of lone-wolf chats. He builds the pattern so your agent stack stops being single-player. Watch this if every teammate is still starting from a blank chat every morning.
Riley Brown — Claude Code Makes Regular Claude Feel Dumb
Riley Brown argues that if you already pay for Claude, Claude Code is the upgrade that makes regular chat feel limited. He demos workflows that run against your whole machine — sites, tools, even 3D and video. Watch this if you are still pasting snippets into a chat box instead of giving the model the filesystem.