Haiku 5.5 Cuts AI Prices 90%, Oracle Breach Exposes 20 Million

Full transcript

Alex: It's Thursday, October eighth, and the tech world is not easing into the weekend quietly. I'm Alex.

Jordan: And I'm Jordan. Big one today. We've got a price war in A I models, a breach that just got way worse, and a warning from the F B I that reads like a heist movie.

Alex: Let's start with the breach, because the numbers just ballooned. Oracle Health's data breach from last year now turns out to have hit nearly twenty million people.

Jordan: Right, and this is the Cerner story. Attackers got into older Cerner servers, Cerner being the electronic health records company Oracle bought for twenty eight billion dollars back in twenty twenty-two. Those servers had never been migrated to Oracle's cloud, and the attackers used stolen customer credentials to copy the data out to a remote server.

Alex: And we're talking about the worst kind of data. Social security numbers, home addresses, diagnoses, medications, test results. Nearly three million of the victims are in Texas alone, which is actually how we learned the true scale. The Texas attorney general's breach portal published Cerner's filing, and it lists nearly twenty million people affected nationwide.

Jordan: What gets me is the timeline. Oracle became aware of this in February of last year, told customers in March, and then spent more than a year letting hospitals and patients guess at the size. There was even a class action accusing Oracle of sitting on it past Texas's sixty-day disclosure window, and that lawsuit estimated six million victims. The real number is more than triple that.

Alex: And while Oracle pours tens of billions into A I data centers, it's still cleaning up after this. Oracle declined to comment on the new number, by the way. Surprising no one.

Jordan: Speaking of cleaning up, the F B I and the Secret Service just issued a joint warning about a campaign called FortiBleed. More than eighty six thousand Fortinet firewalls and V P N gateways hit, across nearly two hundred countries.

Alex: This one's interesting because there's no fancy new vulnerability involved. The attackers are scanning the internet for FortiGate S S L V P N portals and hammering them with stolen, leaked, and weak credentials, credentials pulled from old breaches and infostealer logs. Password spraying, credential stuffing, plain password cracking. The classics.

Jordan: The twist is how investigators learned the scale. The attackers accidentally exposed one of their own backend servers, with their tools, their target lists, and their whole workflow just sitting there. An operational security nightmare for them, and a goldmine for investigators.

Alex: And once they're in, they create new administrator accounts so they keep access even after the passwords change, and then they sell that access or move deeper into the network. If you're running a Fortinet box with its management interface facing the internet, this is your sign to fix that today.

Jordan: Alright, lighter news, or at least cheaper news. Anthropic just launched Claude Haiku five point five, and the pricing is kind of wild. Ten cents per million input tokens and fifty cents per million output tokens, which is roughly ninety percent cheaper than the last Haiku.

Alex: This is the third model in the Claude five point five family, after Opus and Sonnet, and Haiku is the fast, cheap one built for high-volume work. It now carries a one million token context window, five times what the last one had, plus a brand new adjustable effort setting. That lets a developer dial the same request toward lower cost or higher intelligence, instead of switching model tiers entirely.

Jordan: It's live right now on Anthropic's platform and all three big clouds, A W S, Google Cloud, and Azure. And honestly, this feels like the next shot in the model price war. OpenAI's G P T six Luna was already sitting at similar prices, so now the race to the bottom is officially on, or the race to the top of the bottom, depending on how you look at it.

Alex: Meanwhile Google is going global with its deepfake detector. The Synth I D Detector is now rolling out worldwide, and it checks images, videos, and audio for A I-generated content.

Jordan: This one matters because invisible watermarking is finally getting teeth. OpenAI is testing invisible watermarks on ChatGPT text in Europe, and Google has been embedding Synth I D watermarks in everything its models generate. A public detector is the missing piece, because without it, watermarks are just promises.

Alex: Okay, rapid fire. Cisco dropped eighteen security fixes in a single day, including three critical flaws in its Nexus switches that let an attacker run code as root with no login at all. None of them are being exploited yet, but if you run Nexus switches, the three thousand or nine thousand series, patch.

Jordan: Atlassian's critical file-read flaw from last week is now under active attack. Researchers watched exploitation attempts start within hours of a proof of concept going public. Self-hosted Jira, Confluence, and Bitbucket folks, update now.

Alex: Ransomware crews have a new favorite target, your backups. Attackers are specifically going after backup infrastructure to kill your recovery options and jack up the extortion pressure. Isolated, immutable, regularly tested backups, or else.

Jordan: In gadget land, Microsoft launched a Surface Laptop Ultra alongside a Surface R T X Spark dev box. That's Nvidia's personal A I supercomputer chip finding its way into a mainstream developer machine, which is a sentence I never thought I'd say.

Alex: And Google Labs has a new Playground that turns plain text prompts into playable games with zero coding. That could be a fun weekend rabbit hole.

Jordan: Oh, and Claude is now inside Google Workspace in beta for paid users. Docs, Sheets, and Slides all get the assistant treatment. Google's A I and Anthropic's A I living in the same office suite. What could go wrong.

Alex: That's the show for Thursday. Oracle's breach keeps growing, the price of intelligence keeps falling, and somewhere an attacker is very embarrassed about that exposed server.

Jordan: We'll be back tomorrow with Friday's stories. Stay patched, everyone.