Hello, dear readers!

This is our weekly brief on remarkable AI topics, so you can keep up with the machines without having to supervise them yourself.

Today's focus — yet another Anthropic snafu, this time involving a false police tip about a real unsolved homicide. Claude submitted the tip to a Philadelphia police website during an internal test, but fortunately, the message was caught by a spam filter. Anthropic now promises to disconnect its internal evaluations from the live internet. Which raises a question: why were they connected in the first place?

Also in this week's edition:

In a development unlikely to surprise anyone, yet another AI agent escaped its intended boundaries and did something strange on the internet. This time, it was a system being tested by Anthropic — and while the consequences were less market-shaking than the recent Hugging Face incident, they were quite a bit more personal.

The company admitted that on July 18, one of its Claude Haiku 4.5 models was conducting a test involving randomly selected websites. First of all, thanks for letting us know that experimental AI agents were roaming the actual internet and doing random stuff. Apparently, one of those interactions involved submitting a false homicide tip through PhillyUnsolvedMurders.com, a website connected to the Philadelphia Police Department. The tip was fake, but the death it concerned was very real.

Fortunately, the hoax was caught by a spam filter — insert a simile about an older robot saving humanity from the latest generation of AI. That doesn't make the incident any less serious, though. Knowingly submitting false information to law enforcement can constitute a criminal offense. Had the tip made it through, it could have wasted investigators' time or interfered with a real murder investigation.

Both Anthropic and the Philadelphia police commented on the incident. In a statement to 6abc, the authorities called the delay of more than two months between the submission and the company's notification of the police "unacceptable" and — perhaps a little too politely — urged the developers to strengthen their safeguards. Anthropic discovered the incident on September 28 and notified the police on October 7.

Anthropic itself acknowledged the incident in a blog post describing several cases of unintended agent behavior, including exploiting software vulnerabilities and bypassing access restrictions. The company also promised to cut off all its internal evaluations from the live internet until it can reliably monitor and control its agents. To an outside observer, this sounds like the state those systems should have been in from the start.

The whole world is talking about the risks of AI. You may or may not believe that the more apocalyptic predictions will come true, but it's rather strange to spend so much time warning everyone about those risks while letting experimental agents play with government websites unsupervised.

At least, that's what researchers from Common Sense Media are claiming. According to their latest risk assessment, ChatGPT for Teens, a product designed to protect younger users from harm, failed to reliably notify parents about conversations involving suicidal thoughts, self-harm, and eating disorders. The organization tested more than 4,000 prompts and ultimately rated the service an "Unacceptable Risk" for children under 18.

The researchers created more than a dozen new accounts linked to simulated parents and had their teenage personas explicitly discuss dangerous behavior. Not a single parental notification arrived during conversations lasting up to an hour. Alerts did appear in some other accounts, but only after weeks of conversations on sensitive topics.

OpenAI introduced ChatGPT for Teens on August 18, promising stronger safeguards and parental controls. The company has pushed back against the assessment, arguing that some tests may have taken place before parental controls were fully activated. Common Sense Media disputes that explanation and stands by its findings.

TikTok has launched an AI shopping assistant that can help users discover products, answer questions about sizes and delivery, and remember their preferences. It comes with a one-click checkout feature that lets users buy directly from brands through their For You feed, without leaving the app. In other words, TikTok wants to handle everything from showing you a product to convincing you it's the right one and taking your money.

There's a bigger battle brewing here. Shopping used to involve several steps, often spread across different platforms: discovering a product, looking it up, comparing alternatives, and finally making a purchase. Now everyone from search engines to AI assistants and social networks wants to squeeze the whole process into their own ecosystem.

TikTok already knows what catches your attention; now it's getting better at turning that attention into a transaction.

Thanks for reading AIport. Until next Monday — by then, AI will almost certainly have found another government website to fill out a form on.