Good morning! Welcome back to The AI Brief, the daily newsletter from Gizmodo that keeps you up-to-date on the biggest and most bizarre events in AI, with none of the technobabble.

It’s been a big week for cybersecurity, after an OpenAI model helped researchers uncover major vulnerabilities in PHP, a technology on which roughly 70% of the internet is based. Experts are already calling it “Black October.” Meanwhile, banks have their own cybersecurity woes to contend with: customers entrusting AI agents with their finances.

Despite being promoted by tech giants like Meta as reliable assistants, agents have a long and rocky track record of going off the rails in all kinds of unpredictable ways.

In other news, Microsoft is claiming its new AI tool can diagnose diseases more accurately than actual, flesh-and-blood doctors. (It’ll probably be a while before it has a decent bedside manner, though.)

→ Banks get blindsided by AI agents
→ Microsoft says its AI outperforms humans in diagnosing diseases
→ Hugging Face’s new AI model can translate the language of DNA

A litany of leading American banks, including Morgan Stanley and Bank of America, are hastily updating their terms of service in preparation of an onslaught of AI agents—or rather, people allowing those agents to manage their finances (a move we cannot recommend against more strongly).

As Axios first reported on Thursday, some of those banks have already detected instances where agents appear to have acted autonomously within customer accounts, after those customers provided the bots with their login information. (Again, please don’t do this.) The report didn’t provide much in the way of details, and it also noted that banks aren’t able to easily distinguish between agent activity and legitimate human activity. But it seems to be a moment of reckoning for an industry confronted with an army of bots, which many Americans seem all too happy to use—even, apparently, for a task as sensitive as banking.

Luckily for civilization, AI agents can do more than just hack government systems and handle busywork: they can also now diagnose diseases with impressive, indeed superhuman accuracy.

A new AI system developed by Microsoft called Diagnostic Orchestrator functions like a team of digital doctors, analyzing patients’ medical data to figure out what they might be suffering from. In an article published Thursday in the prestigious journal Science, the company said the system “substantially outperformed practicing generalist physicians” in diagnostic accuracy.

Think of it: AI actually being used to save lives, rather than potentially wiping out all of humanity. That could actually be a future worth striving towards.

Hugging Face has developed an AI model designed to make the language of DNA a little more intelligible to humans.

The model, called Carbon-X, was trained on genomic sequences not only from humans, but also from a range of other animals, as well as from plants, fungi, and protozoa, Wired reported on Thursday. It’s able to respond to user queries in natural language, much like an ordinary chatbot like ChatGPT or Claude, and its creators hope that it could lead to the creation of new drugs and genetic therapies. Users could also use it to understand a little more about their own genetic makeup, and whether they could be predisposed to certain illnesses.

None of that personal genetic data—you’ll be happy to learn—gets shared with Hugging Face, a senior research scientist at the company told Wired.

Read more →

The Pentagon is quietly building an elite new tech sabotage agency

Elon Musk makes $1 billion investment in Ellison’s Skydance

This teacherless school says AI can teach students in 2 hours a day

Barclays cutting 5,000 jobs as it increases reliance on AI

Palantir says it cannot shut off NATO AI system as Europeans fear US tech dependence

Claude did not invent scorpion antivenom

‘Know Your Agent’ will be the new rule of the AI age

Microsoft shuts the door on The Garage, the central program behind its hackathon and maker spaces

Recursive launches Flint, a laptop coding model it says was built largely by AI

Gray paint has never had it as good as in the AI boom