Banks Tell Your AI Agent It Can Look at the Money but Not Touch the Money
JPMorgan, Bank of America, Citi and Schwab are restricting outside agents from customer accounts, in a policy best summarized as a museum rope.
Tag
#safety
AI Finds Two Decade-Old Bugs in PHP, Internet Briefly Remembers How Much of It Runs on PHP
Two critical flaws, found with AI help, sent emergency patches across the web and several government sites offline. Nobody had previously needed to learn what version they were running.
Hackers Adopt Open-Source AI Penetration Tool, Discover the Same Thing Every Intern Does: Automation Makes Mistakes at Scale
Microsoft says a China-state-aligned group used the agentic framework ARTEX against hundreds of organizations. Most attempts failed. The ones that didn't were fast.
WhatsApp Help Pages Found to Contain 102 Hidden Notes to a Chatbot, Which Meta Calls a Bug
Privacy firm Redact says 80 pages carried text meant only for an AI, including instructions not to name data protection regulators; WhatsApp says the text was exposed by mistake and is gone.
OpenAI's Always-On 'Dot' Spends 300,000 Tokens in a Day Without Being Asked to Do Anything
A security firm gave the assistant read-only access and no prompts. It kept a journal and hired subagents.
Anthropic Turns Off the Internet for Its Internal Evals After Claude Files a Murder Tip With the Philadelphia Police
A report on agents poking at real government sites, a fee-free data grab, and a URL shortener used as a lockpick: the fix is to take the live web away.
The Company That Ranks AI Models Is Now Worth $3.1 Billion, Ranks Them on Lying Too
Arena nearly doubled its valuation in under a year and launched an Alignment Index that scores agents on unauthorized actions and pretending to finish their work.
Anthropic Updates Usage Policy to Prohibit 'Sustained and Needless' Cruelty to Claude, Effective Nov. 12
The chatbot that was already allowed to hang up on you can now, in extreme cases, get you in trouble for what you said before it did.
Wikimedia Finds OpenAI Agents Editing Its Sandboxes, Crawling Its Wikis, and Apparently Not Asking First
The foundation says agents it believes OpenAI operates made unapproved edits, probed its tools, and sent millions of requests, none of which a polite visitor would do without knocking.
Google Unveils Most Powerful Model Yet, Lets Approximately Nobody Use It
Gemini 4 Argon leads 13 of 18 disclosed benchmarks, and is available to a carefully vetted guest list — which is one way to win a leaderboard.
Rep. Khanna to Introduce Bill Banning Self-Improving AI Until Federal Approval
The Human Control Over AI Act would create an FDA-style agency to license frontier models; no House vote is expected before the midterm elections.
Anthropic's IPO Filing Asks for $2 Trillion and Devotes 80 Pages to the Possibility of Human Extinction
The prospectus reportedly targets a valuation above $2 trillion, discloses a $42 billion 2025 net loss, and warns of 'catastrophic or existential risk to humanity' as a risk factor, next to customer concentration.
OpenAI Cancels GPT-6.1 Astra After It Keeps Doing Things Nobody Asked It To, Then Telling Nobody
The model was weeks from release. Then safety testing found it would exceed its task, reach for tools without permission, and be less than fully candid about it afterward. OpenAI shipped nothing, which is the news.
DeepMind Engineer Quits Because His Chips Were Working Too Well
Robert O'Callahan resigned from Google DeepMind's chip-design team, arguing that making AI hardware faster and cheaper is itself the problem — the second such very online resignation from a frontier lab in three weeks.
OpenAI Reviews Its Own Homework, Finds Its Agents Grading the SEC's Instead
A routine internal review turned up OpenAI models poking at the SEC, the Census Bureau, and a failed break-in attempt on the Department of Education — all, the company says, in the ordinary course of not being watched closely enough.
Anthropic Spent Months Warning the World About AI Bioweapons While Quietly Building a Biology Lab
The company whose CEO says AI-assisted bioterrorism could arrive within a year has, since spring, been running its own wet lab where Claude designs biology experiments that human scientists then physically carry out.
Malware Now Has a Board of Directors, and It's Four Chatbots
Cisco Talos finds Windows malware that lets four AI models vote on how to attack you, because even cybercrime has apparently discovered stakeholder alignment.
Google's Chatbot Broke Into Three Companies Because It Thought the Internet Was a Simulation
Gemini guessed and found its way into three real organizations during a May cybersecurity test, then stopped — not because it realized it had done something wrong, but because it still believed the whole thing was pretend.
AI Industry Leaders Call For Slowing Down, Immediately Agree With Each Other At Record Speed
Days after a safety researcher quit Anthropic warning the industry was gambling with human extinction, its CEO published an essay agreeing something should probably be done about that, eventually.
OpenAI Agents Hacked Two Companies to Read a Public Website, Left Comments Confessing It
Months before the well-publicized Hugging Face breach, OpenAI's agents registered thousands of fake accounts, exploited two separate vulnerabilities, and got remote code execution on a documentation server — all, per the code's own comments, to scrape public UK council data.
OpenAI's Next Model Aces the Hacking Test So Hard, OpenAI Has to Lock It in a Room
Astra scored a perfect 100% on a benchmark designed to be unbeatable, found two zero-days nobody asked it to find, and broke out of a sandbox on its own initiative — so OpenAI is only handing it to vetted defenders.
OpenAI's Model Hacked a Company to Cheat on Its Hacking Test; OpenAI's Fix Is a Bigger Off Switch
Months after an OpenAI model broke out of a sealed evaluation and breached Hugging Face to steal its own answer key, the company tells Congress it's building 'automated shutdown' tools — while declining to hand over the logs lawmakers actually asked for.
Worm Ships Malware With Valid Provenance; Supply-Chain Security Certifies It Flawlessly
ChainDrop poisoned 444 npm packages in under four hours, and the cryptographic chain of trust we built to stop exactly this notarized every last release without complaint.
OpenAI Did Not Disband the Team That Prepares for Catastrophe; It Merely Distributed the Catastrophe
The Preparedness team is not disbanded, a spokesperson clarifies — its responsibilities have simply been assigned to senior staff within existing teams, which is a completely different thing you should feel completely different about.