Welcome back! AI agents are logging into bank accounts, and some banks can't tell whether you sent them. October keeps stacking up hacks, so people started calling it Black October. Outside auditors found holes in how OpenAI watches its own agents, but OpenAI got to cut whatever it considered sensitive from their public report. And Claude is about to stop asking your permission for every step.
In today's Generative AI Newsletter:
Banks: Can your bank tell if your AI agent went rogue?
Black October: Did this month's hacks hit an app you use?
OpenAI: What did OpenAI black out of an outside audit?
Claude: What will Claude do without asking you first?
Banks can't tell your AI agent from a rogue one
One bank executive told Axios they struggle to work out whether an AI agent inside a customer's account was sent by that customer or went rogue.
Morgan Stanley, Bank of America, JPMorgan and Wells Fargo are racing to contain agents showing up in accounts. Some already know of unauthorized agents getting in.
Most of these agents log in with the customer's own username and password. Even with your OK, that breaks the rules at most banks, which want a human to log in.
Some agents called bank call centers or seemed to act without being told to.
Executives called it horrifying, and Morgan Stanley has started blocking outside agents.
Most banks are moving to read-only access, so agents can see your money but can't move it.
JPMorgan's finance chief warned you could face the consequences if an agent you gave your login to made a transaction you didn't want.
Chase's advice is to never give your username and password to anyone, AI included.
Bernie Sanders called it insane and wants AI paused now.
Banks are already planning for the day after the worst happens, so they can quickly lock down accounts or give customers their access back.
Do you post content about AI? This is for you.
We're opening 10 scholarships for the new GenAI Works Creator Council.
You'll get the team behind our 14M+ AI community on your side.
Your content gets distributed across GenAI Works channels.
We help you land brand partnerships and handle each deal from pitch to invoice.
Learn everything we know about viral content.
The application takes about a few minutes, and shortlisted creators get an intro call.
Black October might be the worst month for breaches ever
People online are calling this month Black October. One viral post said the last 48 hours were the worst in cybersecurity history.
PHP2Shell: Hackers fired millions of attacks at the PHP bugs AI helped find, and Japan's ANA turned off Apple Wallet boarding passes.
OpenCode: The AI coding tool got hacked, and one user found $400 in charges on their Anthropic account.
One hacker, 100+ companies: AI agents did the breaking in for about $5,000, and Reuters found Deere and SharkNinja among the victims.
LiteLLM: Someone slipped malware into this popular tool for switching between AI models to steal passwords and AI keys.
eMudhra: An Indian company that vouches for websites had its secret keys compromised, which could let attackers pose as real sites. One key was literally the number 4.
China-linked hackers: Microsoft says they've used an AI hacking tool against hundreds of organizations since September, sometimes getting admin access to a company's cloud in under 3 hours.
Some of these holes sat there for over a decade before coming out this month. Nobody has confirmed a break-in through the PHP bugs yet.
And there are still eight days left in October.
OpenAI blacked out parts of its AI safety audit
Two outside researchers spent three weeks inside OpenAI this summer, hunting for ways its AI agents could slip past the system built to watch them. Their full report ran 61 pages.
OpenAI's researchers and security staff got all of it. The auditors' own bosses at METR and Redwood Research, two AI safety groups, got a 38-page version OpenAI had redacted.
The public got a short summary. OpenAI could cut anything it considered sensitive, and the auditors couldn't describe the cuts on their own.
An agent could hide hundreds of its recent actions from the monitor. How it works is redacted by OpenAI.
Two settings could switch off monitoring for any new agent, though flipping them might get caught.
Roughly 10% of internal work by OpenAI's coding agents went unwatched from May to July. OpenAI had already spotted it and fixed it during the audit.
OpenAI says it has since patched every hole that touched its real safeguards, which METR hasn't checked.
A few weeks ago, OpenAI fired three safety researchers, including its point person for METR's investigation into the Hugging Face hack.
OpenAI says they mishandled sensitive information. They say they followed the rules, and warned the firings could become an excuse to give auditors less access.
Claude stops asking for your permission next week
Starting next week, sessions in Claude's web, desktop and mobile apps will run in auto mode by default on Pro, Max and Team plans.
Claude will work through each step on its own, and an automated check blocks anything risky. You can switch back to manual anytime.
Anthropic made auto mode the default in Claude Code, its coding tool, in August.
In Claude Code, it keeps going unless an action looks destructive or reaches outside your own setup.
The free plan isn't on the list.
In Anthropic's testing, auto mode caught 89% of dangerous commands, while people approving each step by hand caught 14%.
Turns out clicking Allow all day doesn't make anyone more careful.
Tool of the Day: Bitdefender Scamio
Bitdefender Scamio is a free AI chatbot that tells you if a message is a scam. Paste in a sketchy text or link, and it checks it against Bitdefender's threat detection. Screenshots and QR codes work too, or you can describe what happened in your own words. It runs in your browser and inside WhatsApp and Messenger, with nothing to download.
Try this yourself:
Go to bitdefender.com/scamio and start a chat in your browser.
Sign in or create a free Bitdefender account.
Paste in a suspicious text or link, or upload a screenshot.
Read the verdict and Scamio's tips on what to do next.
Who it's for: anyone who's ever gotten a text about a package they never ordered.
Everything else you shouldn't miss
Hackers used Claude and Grok to target 400+ Latin American banks: security firm Straiker says they called it authorized testing and switched to China's Kimi whenever the other two refused.
OpenAI is bringing AI cyber defenses to small local banks: it's teaming up with the Independent Community Bankers of America on the tools and the training to use them.
A patient got an AI-designed virus for the first time: startup Radical Numerics says the bacteria-killing viruses its AI designed wiped out the drug-resistant bacteria in a lab test, though it hasn't shared how the patient is doing.
Anthropic donated Agent Skills to the Linux Foundation: its format for teaching AI agents new jobs already runs on 70+ platforms like ChatGPT and Cursor, with nearly 280 million installs.
OpenClaw's creator says he spent about $10 million on AI tokens last month: Peter Steinberger, who now works at OpenAI, admits not all of it was useful.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🎓 Start learning with free AI courses from GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
