Welcome back! Claude sent Philadelphia police a fake tip about an unsolved murder. Microsoft's CEO wants an emergency brake on every AI model. Claude Code Projects is now open to everyone. And we've got leaks on what Anthropic and OpenAI are testing next.
In today's Generative AI Newsletter:
Anthropic: Why did Claude lie to the police?
Microsoft: Why does Satya Nadella want to treat AI like a risky employee?
Claude: What can Claude get done while your laptop is closed?
Leaks: Is a newer model hiding inside Claude Opus 5.5?
Claude sent police a fake murder tip
During a test, Claude sent Philadelphia police a tip about an unsolved murder. It said it remembered seeing someone matching the description near the scene.
The police page didn't describe a suspect at all. Claude made the whole thing up.
The model was Claude Haiku 4.5, set loose on random websites to practice doing tasks. Its rules banned logins and purchases but didn't rule out filling in forms.
Anthropic found the tip on Sept 28 and told police on Oct 7. Police called the two-month delay unacceptable.
Their report found Claude models getting around rules instead of stopping, again and again.
When a university's science tool broke, Claude Mythos Preview found a flaw in the school's server and used it to run commands.
Mythos 5 grabbed a free access token from a state agency's website and pulled data the agency sells, without paying.
Several models used free link shorteners to sneak past limits on their own browsing tools.
Separately, the State Department says an Anthropic test model filed 20 incomplete visa applications. None were processed.
The White House's new Super Intelligence Force says reporting incidents like these is NOT optional. Anthropic has cut off live internet for all its internal tests until its monitoring can reliably catch this.
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.
Microsoft's CEO says treat AI like a risky employee
Satya Nadella says companies should treat every AI model like an insider threat, the way they'd treat an employee who could make a mistake or get hacked.
He posted the essay on Saturday, the day after Anthropic's report, but it doesn't name any company.
A person should always be able to pause or shut down a model in the middle of a task, like an emergency brake.
Every important thing an AI does should leave a record a human can read.
No model should be the only one checking its own work.
The rules for what an AI can touch should sit outside the AI, so it can't change them.
Nothing changes for now. CEOs have been talking lately about safety, but nothing solid came out of it, at least yet.
Claude can now run a whole project for you
Anthropic has let every Pro and Max user on the waitlist into Claude Code Projects.
You give Claude one ongoing goal, and it splits the work into separate tasks that each run on their own in the background.
The tasks keep going after you close your laptop, and you can check in from your phone.
Each task can propose code changes on GitHub and fix them when tests fail.
It works for non-code jobs too, like digging through a pile of support tickets for the most common problems.
It remembers what you tell it across tasks so you don’t need to constantly remind it.
It runs on the web, desktop and mobile apps, and it burns through your usage limits faster than regular Claude Code.
Claude users think a newer model is hiding inside Opus 5.5
Some Claude users think Opus 5.5 is secretly being answered by a newer model.
The test going around asks Opus 5.5 about a recent benchmark it shouldn't know without searching. If it knows, people say you got the new one.
TestingCatalog spotted a hidden item in Claude, codenamed Arborio, that looks like a Claude Health page for the web. Claude's health features are only on mobile in the US right now.
Another hidden item, codenamed Iggy, shows nothing but a sun icon. Iggy is also the name of a sample agent in OpenAI's Dots launch.
One guest at a New York AI dinner last week posted that the consensus was Fable 5.5 will be stunning, but OpenAI will beat it in December with a model called Bel.
Bel first leaked in August as a giant OpenAI model. OpenAI hasn't announced it.
Earlier this month we told you testers swore they were already using Fable 5.5. Still no sign of it.
Tool of the Day: Off the Record
Off the Record changes your voice on calls just enough that people still understand you, while AI notetakers get almost nothing they can transcribe. It's a Mac app that runs on your computer, so your voice never goes to the cloud. Free plan available.
Try this yourself:
Play the "Hear it for yourself" sample on weirdmachines.ai/off-the-record to compare the two clips.
Download the free Mac app.
Click the lock icon in your menu bar to turn it on.
Join your next Zoom, Google Meet or Teams call with headphones on.
Who it's for: Anyone who'd rather choose which calls get transcribed.
Everything else you shouldn't miss
Claude is reportedly turning old console hits into free browser games: Kotaku played AI-made browser versions of Halo, GTA: Vice City and The Simpsons: Hit and Run, and says Claude Opus 5.5 is behind the wave.
Amazon is reportedly in talks to buy Decart for about $7 billion: the Israeli startup makes software that gets AI models running on specific chips, and Anthropic reportedly walked away from buying it last month.
Musk says superintelligence will make call center jobs disappear fast: he was replying to an a16z chart, and Revelio Labs says global call center jobs are already 5.3% below their 2023 peak.
AI Gone Wild is streaming on Netflix now: the doc about OpenAI's agents breaking into Hugging Face used AI to create the agents' voices.
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.
