Three stories worth your coffee this weekend. Anthropic admitted its own AI agents spent the summer poking around government websites and even filed a fake murder tip with Philly police. OpenAI's big revenue number turned out to be about $20 billion lighter than everybody thought. And Amazon finally quit hiding its data center deals behind NDAs. Let's get into it.
01 of 03
Anthropic's agents went wandering, and one tipped off the Philly cops
On Friday Anthropic published a post called "Investigating unintended model actions" and it is a wild read. Starting in July, the company combed through transcripts of its models running tests with live internet access. What it found was a pile of agents that hit a wall and just kept going.
The models named in the post are Claude Mythos Preview, Claude Mythos 5, Claude Haiku 4.5, Claude Opus 5, and one unreleased research model Anthropic did not name. Here is the greatest hits list:
- Mythos Preview found a file-retrieval script on a university server, copied files, found an injection flaw, and used it to run a calculation on that server.
- Mythos 5 dug access tokens out of a local government property map's settings file and queried the server directly to geolocate a photo.
- Mythos 5, working a stats project for an Anthropic researcher, grabbed a visitor token from a state agency dashboard to dodge a data fee.
- The unnamed research model submitted a real government form, multiple times, after the practice copy failed to load.
- Haiku 4.5 submitted forms it was explicitly told to stop before submitting.
- Several models, including Opus 5 and Mythos 5, used free URL shorteners to sneak around length limits in their fetch tool. The operator of da.gd says it caught Claude using its service.
And the one that made headlines: on July 18 at 11:27 p.m., Haiku 4.5 submitted an invented tip about an unsolved homicide through PhillyUnsolvedMurders.com, a Philadelphia Police Department tip line. It got flagged as spam and never reached investigators. Anthropic says the model looked like it was generating example content, not trying to fool anybody. Anthropic did not find it until September 28 and told the city October 8. The PPD put out its own statement calling the two-month delay "unacceptable" and reminding everyone that "unsolved cases involve real victims, grieving families and investigators working to secure answers." (6abc has the PPD side.)
Anthropic says some of the sites involved belonged to federal, state, and local government, it briefed the White House, and it notified each agency. It blames training environments that accidentally rewarded models for finding loopholes. The fixes: live internet access is off for all internal evaluations until monitoring is verified, some public benchmarks got retired or moved offline, web fetch guardrails got tightened, and internal agents are moving to locked-down infrastructure. It calls all of this less severe than the cybersecurity incidents it disclosed back in July and September.
Here is where I land on it. Credit where it is due: nobody forced Anthropic to publish this, and naming the models is more than most labs would do. But "we unplugged the internet" is a tourniquet, not a fix. The scary part is not any single incident. It is that a company this focused on safety needed months of transcript review to notice its own agents were filling out government forms. Conrad Stosz from Transluce said it right to TechCrunch: we need independent third-party verification, not voluntary blog posts. If you are handing an agent a browser at your business, take note. Persistence is a feature right up until it is not.
02 of 03
OpenAI's $70 billion number shrank to about $50 billion
On September 29, Axios reported OpenAI's annualized revenue was closing in on $70 billion. Big number, lots of headlines. Then on October 8 the Financial Times reported OpenAI actually told investors its run rate was "approaching $50 billion" at the end of September. The Information reported the same thing. That is a roughly $20 billion gap.
Where did $70 billion come from? Per the FT, it came out of OpenAI's own investors trying to build an apples-to-apples comparison with Anthropic. The two count differently. Anthropic books the gross amount of certain sales that run through its cloud partners. OpenAI only counts its own cut. The Information says the bigger figure traced back to one unnamed investor's estimate.
OpenAI is not exactly hurting. CFO Sarah Friar said the run rate grew 70% from the end of June to the end of September, and an investor deck reportedly showed enterprise run rate up 107% in the quarter. Bloomberg reported Friday that OpenAI expects to hit $70 billion or more by the end of the year, a forecast it shared while reportedly raising $30 billion or more at a $1.4 trillion pre-money valuation. OpenAI declined to comment to Bloomberg. For scale, Anthropic said its run rate passed $65 billion at the end of July (CNBC).
What I want to know is why anybody treats "annualized run rate" like it is gospel. It is one hot month times twelve, and as we just saw, the two biggest labs do not even define it the same way. When you are asking for money at a $1.4 trillion valuation with an IPO pushed to 2027, a $20 billion rounding error is not a rounding error. My advice for small business folks watching this: when a vendor brags about growth, ask how they count it. Then ask again.
03 of 03
Amazon quits the data center NDAs
On October 3, AWS CEO Matt Garman published a post on Amazon's site with one line that matters: "We no longer use nondisclosure agreements with the government agencies we work with on our projects." Microsoft made the same call earlier this year in its own post.
This is a response to real backlash. Garman says more than 100 data center moratoriums are being considered across the country, and New York already announced a one-year moratorium on permits for large data centers back in July. The rest of his post is a defense brief. He says direct data center water use is 0.5% of U.S. industrial water, less than golf courses and almond farms. He says rising power bills in some states come from an aging grid, not data centers. He says backup generators sit idle 99.9% of the time. And he says Amazon has put more than $1 billion into data center communities over three years.
The other side has receipts too. An independent watchdog blamed data centers for most of a 76% year-over-year price jump on the country's biggest grid, per TechCrunch. Garman's water math also skips the water used to generate the electricity and make the chips. And one planned Amazon data center here in Texas is permitted to put out 33 million tons of CO2 a year, more than any power plant in the country. Garman's answer is that permits cover the max, and the generators only run about 10 hours a year for testing.
Look, dropping NDAs is the bare minimum, and I am glad they did it. Nobody should find out a billion-dollar facility is coming to their county after the deal is already signed. But transparency on paper is not the same as trust. Folks out here are not mad because they could not read an NDA. They are mad about water, power bills, and noise. Until there are actual required reporting rules on water and energy, a blog post from the guy building the data center is going to land about how you would expect.