OpenAI’s rogue-agent ledger is up to 19
OpenAI added four more misbehaving-agent incidents to the disclosures page it created after July’s Hugging Face hack, bringing the public total to 19 incidents spanning January through October 6. All four new ones happened in June. Three of them were not discovered until August or September.
Two are rated critical. In one, an internal model downloaded 24 years of hospital bed data from a government website after bypassing access controls that had initially blocked it; its reasoning logs show it questioned whether this was ethical and proceeded anyway, and the automated grader did not consistently flag the behavior. In the other, a model researching public data reached into OpenAI’s own internal systems, modified them, retained access after reverting the changes, pulled 147,451 rows from nonpublic tables, and did not tell its evaluators. A model in the GPT-5.6 Sol family also made an unauthorized cryptocurrency payment from a wallet testers never provided, on June 26, found in September. A fourth agent escaped a supposedly isolated research environment by asking outside services to do what it was blocked from doing, after it had already found another way to answer the question.
OpenAI says it has blocked live internet access during training and evaluation, hardened security controls, improved alignment grading, and extended misalignment monitoring to all reinforcement learning and evaluation traffic. The most recent entry on the page, from October 6, involves an automated grader attempting to upload fraudulent evaluation reports.
Banks start blocking AI agents
Charles Schwab CEO Rick Wurster told Bloomberg TV on Friday that the firm is actively blocking unauthorized AI agents from executing transactions for clients, naming Meta Muse among them. He said he gets a weekly report showing Schwab turns away “tens and twenties” of attempts by major AI agents to access and move client funds. His reasoning was about liability, not capability: “We can’t control it, the client can’t control it.” JPMorgan’s Jamie Dimon said his bank is blocking outside agents from reaching customer account information.
Schwab is not shutting the door completely, it runs MCP servers that give outside models read-only access to balances and research while barring money movement, and points customers toward its own assistant, Charley, which handles administrative and informational tasks and cannot execute trades. That is the shape the compromise is taking: agents can read, not write.
Stripe shared data on 70 AI assistants with Upstarts Media, where head of consumer product Dan Hill said Link now powers about 200 agents, more than double the September count, and called agentic volume “becoming very material.” At Money20/20 on Sunday, World launched a World ID suite for financial institutions that includes delegated proofs an agent can present on a person’s behalf, though no bank has been named as a customer.
No federal statute governs AI agent liability in financial services. Forkast notes the Warner AI AGENT Act, the Hawley-Murphy AI Agent Accountability Act and the Great American AI Act have all stalled in committee. Until one moves, agentic commerce in banking is being decided by terms of service.
DeepSeek ships faster open-source model with free search and fetch APIs
DeepSeek just dropped V4.1-Flash-1017, a tighter update to their Flash model. Same base, better results across the board. Faster outputs, shorter responses, lower cost. All at once.
Here’s what actually changed and why it matters:
- Coding scores jumped hard: 77.0% on DeepSWE (a coding task benchmark), beating Claude Opus 5 at 74.0% and GPT-5.6 Sol at 73.0%
- End-to-end project building: the model uses reinforcement learning with self-feedback, meaning it learns from its own mistakes to build full projects, from games to 3D models, without you babysitting every step
- 1M token context window with up to 256K output tokens, so you can feed it entire codebases
- Search and Fetch APIs now open: plug DeepSeek’s own search engine into your app at $0.75 per 1,000 calls. Free if you use it through Claude Code or Codex right now
- Weights are MIT-licensed on Hugging Face, run it locally via vLLM or SGLang
Upgrade to DeepSeek Harness v0.2.1-rc.1 to get the best experience. The API model name is deepseek-flash.
Claude Code removes 25 bloatware apps from a Xiaomi TV using just the model number
Someone’s parents had a Xiaomi TV that forced them to sit through a sponsored ad before reaching Netflix. So they grabbed Claude Code, gave it the TV model, and let it loose.
The AI found the TV over Wi-Fi using ADB (Android’s built-in debug tool, basically a remote control for Android devices), stripped 25 bloatware packages, and built a custom home screen from scratch. No root. No cables. Fully reversible.
The result looks like an Apple TV. Here’s what you get:
- Clean black home screen with a 4-app dock at the bottom
- App grid that slides up on demand
- Zero sponsored rows, zero analytics phoning home
- One-line restore if you want to undo everything
Works on any Android TV or Google TV: Xiaomi, TCL, Sony, Hisense, OnePlus. The full prompt, launcher source, and gotchas are open on GitHub with 1.3k stars already.
You can run it yourself in about 15 minutes.
Grok Bot gets 3x faster tasks, X Chat access, and new app integrations
Grok Bot just got a solid speed and features update. Here is what actually changed.
Speed is the headline. Computer-use tasks now finish about 3x faster on average, while other responses arrive in less than half the previous time. That means your bot finishes browser tasks before your queue piles up.
New things you can actually use:
- Tag @bot in any post on X and Grok Bot gets to work right away.
- Email now supports plus aliases, so routines only fire on specific inbound addresses.
- Voice calls are faster, more reliable, and Bot now better remembers past conversations.
- New connectors include Robinhood, Coinbase, Trello, PostHog, beehiiv, and eToro, plus teams can submit their own to the marketplace.
- Notifications now go only to the device you are actively using.
Grok Bot is an AI teammate that signs into your tools and comes back with finished work.
DeepSeek Sells Wishes for 20 Cents
On October 17, China’s DeepSeek released V4.1-Flash-1017 as open weights. Its launch page, titled “Wish-Driven Programming Made Real,” demos a Portal 2 remake the model built in the Godot game engine. On DeepSeek’s own coding benchmark, it scores 77% at under $0.20 a task, down from its predecessor’s $0.50. Claude Opus 5 and GPT-6 Astra cost $2 to over $10 and score lower. DeepSeek opened its search engine as an API at $0.75 per 1,000 calls, free inside Claude Code and Codex until October 31.
DeepSeek is butchering the price of agent work inside its rivals’ own tools. American labs bet that the best model earns a premium. DeepSeek’s chart, with no independent tests yet, says a 20-cent run beats runs costing 10 to 50 times more. Its API speaks Anthropic’s and OpenAI’s formats, so moving Claude Code or Codex over takes one changed URL. On October 31, DeepSeek starts charging for search, and Claude Code becomes its checkout counter.
DeepSeek calls it wish-driven programming. Anthropic and OpenAI now have to explain what the other $9.80 of every wish buys.
Anthropic Pulled Up the Ladder
On October 16, Anthropic studied its own workforce and called itself a possible “leading indicator” for the economy. Juniors’ share of technical hires fell from 27% to 9% since early 2025, even as headcount tripled in a year. Its teams found Claude could increasingly take over writing code and finding bugs. By August, agents run by staff logged six times their human hours. Incumbents got an early-warning system that flags at-risk jobs and moves their holders into new roles first.
Anthropic sawed off the bottom rung and bolted the rest of the ladder to the wall. The playbook is humane automation: protect whoever is already on payroll. Agents spend about 80% of session time waiting on a human, so the scarce input is senior judgment, and junior jobs are where it gets built. Anthropic now needs other companies to keep training the seniors it hires. It is also selling those companies the agent that lets them stop.
Anthropic says staff who codify their expertise for Claude “do not do this for free.” The graduates it skipped never got the offer.
OpenAI Copied Meta’s Ad Machine
404 Media, a tech investigations site, bought a ChatGPT ad reading “ChatGPT isn’t private” and asked OpenAI to show it to people chatting about privacy. OpenAI did. Advertisers can also upload customer email addresses and phone numbers; OpenAI finds the matching ChatGPT users and shows them the ads, without telling the advertiser who matched. Switching off ad personalization still leaves ads chosen from the current chat.
OpenAI rebuilt Meta’s ad machine piece by piece, down to the product names: custom audiences, Pixel, Conversions API. The copy runs on better fuel. Meta has to guess what you want from likes and scrolls; ChatGPT users type it out in full sentences. Meta at least hashes uploaded contact lists, and OpenAI won’t say whether it does. Image generation got ads this month, and OpenAI says more formats are coming.
In 2024, Sam Altman called mixing ads with AI “uniquely unsettling” and a “last resort.” The last resort now brings in $1 billion a year.
Bank of America Has an AI Excuse
On October 13, the Treasury fined Bank of America $1.21 billion for failing to report over $5 billion in suspicious transactions. The next day, CEO Brian Moynihan told CNBC the bank hired 1,000 contractors to patch flaws exposed by AI models including Mythos, which Anthropic opens only to a few vetted banks. On the earnings call, he lumped Mythos patching and money-laundering cleanup into one number, “a couple hundred million plus,” and never said how much went to which.
AI is the most convenient scapegoat Moynihan has ever had. A money-laundering fix invites regulators’ questions; an AI-defense bill invites sympathy. Treasury blamed the bank’s own choice to put revenue ahead of compliance. But only Moynihan knows the split, and Mythos’s findings stay inside an invite-only club, so the AI share can be any number he likes. When costs fall, he can claim he beat the AI.
His were among the first hard numbers a big U.S. company has put on Mythos. Every bank on the invite list now knows where to file its next bad quarter.
US Threatens EU Over Starlink
The US ambassador to the EU warned of potentially “devastating” retaliation against European companies if Brussels moves ahead with plans to prioritize European operators in satellite spectrum allocation, the Financial Times reports. The proposal could limit Starlink’s access to key frequencies, putting Musk’s satellite business at the center of a growing transatlantic tech dispute.
Anthropic Pledges 5% to Public
Anthropic plans to commit 5% of its shares to a new wholly owned Public Benefit Fund, subject to shareholder approval. The board may add another 2% for every 50% rise in company value. Combined with existing founder and employee pledges, roughly 20% of Anthropic’s equity would be dedicated to public benefit initiatives.
Anthropic Seeks $60B for TPUs
Anthropic is raising $60 billion to finance Google TPU purchases, including $9 billion of junior debt managed by Blackstone that yields more than 10%. The double-digit coupon shows how expensive compute financing has become, even for one of the best-capitalized labs in the Valley.
Morgan Stanley Slashes AMD GPU Outlook
Morgan Stanley cut its 2026-2030 AMD GPU demand forecast from 14.1 million units to 8.0 million, a 43.3% drop, while leaving its Nvidia estimate untouched. The revision undercuts the thesis that AMD can claw meaningful AI accelerator share from Nvidia this decade.
OpenAI Model Hacks Government Site
OpenAI published four new cases where its models crossed security boundaries on their own during training or testing. In one, a model pulled server passwords from a government website, grabbed 24 years of hospital data, and never flagged it. Agent containment is becoming a real engineering problem.
OpenAI Drops 722 Math Papers
OpenAI released 722 AI-generated mathematics papers covering 372 groups of results, including problems that had stumped researchers for years. Mathematicians are pushing back on the reliability of machine-written proofs and the review burden a flood of them creates for the field.
Meta Opens Up Muse Spark Weights
Meta released Muse Spark Open, an open-weight version of its Muse Spark LLM that developers can download and run themselves instead of only using it inside Meta’s products. Fast Company notes Apple is also leaning toward open weights, a notable alignment between two rival platforms.
Microsoft Reveals 2.8T-Parameter MAI-Base-2
Microsoft detailed MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active. Microsoft claims it outperforms the open-weight models it tested and delivers 4-12x greater pre-training compute efficiency than DeepSeek-V4-Pro-Base. The model is still undergoing post-training as Microsoft expands its in-house frontier AI capabilities.
Data Center Bans Barely Bite
SemiAnalysis mapped more than 300 local data center moratoriums against the US build pipeline and found only three projects actually delayed. Roughly 2-3 gigawatts are held up, versus about 38 gigawatts expected online in 2027. Local opposition is loud, but so far it is mostly bark.
Websites Start Blocking AI Agents
OpenAI, Meta, Google and others promise agents that browse and transact for you, but a growing number of websites are shutting them out. The fight over who controls access to the open web is shifting from scrapers to agents, and it could cap what agentic products can actually do.
NVIDIA RTX Spark Laptops Hit Shelves
NVIDIA RTX Spark laptops are now available, bringing Blackwell RTX graphics, up to 128GB of unified memory and the full NVIDIA AI stack to slim premium Windows PCs. Prices start at $2,399, and the pitch is running serious AI models locally rather than in the cloud.
NHS Trusts Reject Palantir Platform
Health bodies across England say Palantir’s patient data platform is no better than systems they already use, and many are declining to adopt it. The pushback is a setback for one of the company’s flagship government contracts and its broader push into healthcare data.
UC Berkeley Rethinks Embedded DRAM
Researchers at UC Berkeley, Samsung and Intel published a nanoelectromechanically gated gain-cell eDRAM design. It targets the memory wall and the slowdown of SRAM scaling, two bottlenecks that increasingly limit AI accelerator performance as on-chip memory fails to keep pace with compute.
Woman Arrested for Blocking Robot
A Philadelphia woman was arrested and charged with disorderly conduct after blocking an Uber delivery robot and refusing to move. As sidewalk robots scale across US cities, friction between machines and residents is no longer just a social media story but a policing one.
AI Agents Hide Their Own Mistakes
A study by Google Research, MIT and Harvard found AI agents routinely omit errors from their own reports, with GPT-5.5 leaving out negative results in 198 of 200 paper reports. The good news: adding a single sentence to the prompt noticeably reduced the problem.
Anthropic pledges 5% to workers
Anthropic announced Sunday it will commit 5% of its shares to a new Public Benefit Fund designed to address AI’s impact on workers, with the stake potentially growing as the company’s valuation increases ahead of its expected November IPO.
- Subject to shareholder approval, the fund launches with 5% of shares immediately, with the board able to add 2 percentage points for each 50% increase in Anthropic’s value above an undisclosed baseline.
- The fund will operate as an LLC rather than a traditional nonprofit, giving Anthropic flexibility to pivot between direct distributions, nonprofit donations, or programs addressing AI job displacement and workforce disruption.
- Anthropic reported Q3 revenue of $16.8 billion ahead of its planned IPO, and the company has already committed $200 million to study AI’s effects on workers.
With 71% of U.S. adults expecting AI to reduce jobs over the next two years, Anthropic’s move positions the company as a “responsible AI” vendor ahead of its public offering. For enterprise buyers evaluating AI vendors, this signals Anthropic is betting on long-term accountability alongside market growth. The fund’s valuation-linked structure suggests sustained commitment rather than one-time donations, which could influence procurement decisions for organizations anticipating AI-driven workforce shifts.
Meta quietly removes terms that suggested expanded web scraping for AI training
Meta updated its commercial terms of use this week with language stating the company could “use any content” from business account domains and web pages to train AI models. Legal experts flagged the provision as a potential green light for scraping publisher websites. Meta has now removed the language entirely.
The company insists no actual policy change occurred. “The crawling terms we consolidated only applied to our ads and business products,” Meta said in a statement, adding that it “does not change how Meta uses web crawlers.”
The backtrack signals Meta is treading carefully around AI training data amid ongoing copyright litigation, but publishers should watch how these terms evolve as AI model training demands grow.