Transcript
Kate: Good morning! It's Tuesday, October twenty-seventh, and this is your daily update on security, AI and tech. Dade, security first, because there's a lot of patching going on.
Dade: Morning, Kate. Yes, and the headline is PHP. CISA added two PHP vulnerabilities to its Known Exploited Vulnerabilities catalog yesterday. One is a memory corruption bug, the other a stale pointer issue, and CISA updated its PHP alert saying critical flaws need immediate patching. That means they're being exploited in the wild, not just theoretical.
Kate: And there's a second PHP story on top of that, right?
Dade: Right. PHP rebuilt its official Windows packages for versions eight point two through eight point five, the second Windows update in five days. This one fixes a flaw in the Brotli compression library that's bundled in those packages. It can corrupt memory and potentially allow code execution. And Microsoft has pushed emergency dot net fixes for the same Google Brotli bug. So if you run PHP on Windows, or dot net, check your versions today. For a consultant, this is a nice example of a bundled dependency turning into your problem.
Kate: Also worth flagging for Belgian listeners: the Belgian Centre for Cybersecurity warns about a remote code execution flaw in ISC BIND nine. Patch immediately, they say.
Dade: Yes, and Fortinet adds to the pile. They've shipped hardening updates for FortiOS, FortiManager, FortiAnalyzer, FortiMail and FortiAuthenticator, but they're keeping the vulnerability details secret until November tenth. The flaws were reportedly found with AI help. My advice: don't wait for the details. Patch now, because once the advisory lands, attackers will diff the patches fast.
Kate: Let's move to identity attacks. Huntress describes a Microsoft 365 campaign using token replay.
Dade: Exactly. Attackers use an adversary-in-the-middle relay to capture the sign-in session, then replay the stolen token to get into email. Ninety-seven accounts at seventy-nine organizations. The point is that multi-factor authentication didn't stop it, because the token is already past the MFA step. So conditional access, token protection and short session lifetimes matter. And related: researchers see hundreds of cases where attackers use Microsoft Teams external chat to pose as IT support, delivering malware and even ransomware. Check whether your tenant allows external chat from anyone.
Kate: So, same old trick, new channel. Classic help desk impersonation.
Dade: Yeah, exactly. People trust Teams more than email.
Kate: Next theme: supply chain, and AI agents. Dade, what's new?
Dade: Two things. First, a campaign called Gheist is abusing stolen GitHub tokens to harvest cloud secrets from thousands of developer repositories. Second, the LiteLLM supply chain attack apparently started with an exposed master key that was never rotated. Different incidents, same lesson: if a secret might have been exposed, rotate it. Not later, now. There's also a worm that armed forty-two GitHub repositories with payloads around thirteen megabytes, big enough to dodge GitHub's code search.
Kate: That's sneaky. And the AI agent angle?
Dade: This one's interesting. Researchers at Forever Security found that a file called AGENTS.md, dropped into the Downloads folder by a website without approval, can trick AI coding agents opened in that folder into installing a malicious npm package or handing over a login token. Dark Reading also reports an in-the-wild attack using the same idea, where agents trust local instruction files in markdown. The line between a text document and an executable action is blurring. Treat those instruction files like code.
Kate: Which ties into a survey I saw: security teams worry more about employees pasting company data into AI tools than about AI-powered attacks.
Dade: Yeah, the biggest AI risk comes with an employee badge. Makes sense.
Kate: Now AI. OpenAI released GPT-six point one Terra.
Dade: It's a fast mid-tier model that emphasizes visual reasoning and operating desktop software. API prices are fifty percent below GPT-five point six Terra. The security twist: Help Net Security notes it meets OpenAI's own Critical cyber threshold. So the cheaper it gets, the more capability is widely available. Meanwhile Texas has sued OpenAI for deceiving consumers about safety, and it wants Sam Altman held personally liable. That's unusual.
Kate: And Anthropic filed its IPO prospectus.
Dade: Yes, three hundred thirty-five pages, about a hundred on risk. The number that jumped out is about eight hundred billion dollars in planned AI spending, and the filing makes the dependence on gas for power pretty visible. Markets liked it: the Nasdaq rose almost one percent on the optimism. Separately, Anthropic published the safeguards behind Claude Duet, its proactive agent, including approvals and checks for sensitive information.
Kate: Last theme, tech. Microsoft is offering voluntary buyouts?
Dade: To around sixty thousand employees, roughly a quarter of the workforce, in the US, Canada, India and the UK. Voluntary, with a year-end deadline. And a quick one for cloud people: AWS says it successfully tested a sovereign European cloud that can run fully independently inside the EU. Whether customers believe that is another question.
Kate: And Apple?
Dade: iOS twenty-seven point two is out, fixing more than a hundred and sixty vulnerabilities, so update. It also adds hidden SynthID watermarks to AI images. A new iPad mini arrives, and M6 iMacs and MacBook Pros are expected today. And back in the Netherlands, police opened a criminal investigation into the cyberattack that hit Albert Heijn and other Ahold Delhize chains.
Kate: That's the briefing. Thanks, Dade. Patch PHP, rotate your secrets, and have a great day, everyone!
Dade: Have a good one, and stay safe out there.
No news overview was kept for this day.