Threat Vectr Weekly — week of Jul 27

Welcome to Threat Vectr Weekly for the week of July twenty-seventh. I'm Marcus, and we have a packed episode for you today. An AI tool built to defend systems is raising questions about who else might use it. A critical flaw in enterprise software that runs workflows for most of the Fortune 500 is already being exploited. WordPress sites by the millions need an update right now. And in a genuinely strange first, Hugging Face says it was broken into by another AI. All of that, plus election claims, insider threats, and a supply chain attack hiding in plain sight on a trusted developer platform. Let's get into it.

We start with something that sounds like science fiction but is very much real. Anthropic, the company behind the Claude family of AI models, quietly released a specialized version called Claude Mythos back in April of twenty twenty-three. It went to a small group of fifty technology partners under a program called Project Glasswing. The goal was to give defenders a head start, using AI to find security flaws faster than attackers could exploit them. And it worked. Those fifty partners uncovered more than ten thousand critical vulnerabilities across major operating systems and web browsers, including a bug that had been hiding undetected in OpenBSD, a widely used operating system, for twenty-seven years. But here is the tension. A tool powerful enough to find ten thousand flaws in months is also powerful enough to help someone cause harm. Anthropic restricted access for exactly that reason, and by June of twenty twenty-three, the U.S. government had imposed export controls on Mythos. The takeaway: AI is becoming a genuine force multiplier for defense, but access controls and oversight are not optional extras. They are part of the product.

That dual-use problem is one we will keep coming back to. Speaking of things that move faster than comfortable, let's talk about ServiceNow. If you work in a large company, there is a very good chance your internal IT requests, HR workflows, and help desk tickets all run through ServiceNow's AI Platform. Eighty-five percent of Fortune 500 companies use it, and the platform handles more than one hundred billion workflows a year. That scale is exactly why a vulnerability tracked as CVE-2026-6875 is so alarming. The bug requires no username, no password, and no special access. An attacker from anywhere on the internet can reach a sensitive part of the platform and run their own code on the server. ServiceNow patched its hosted instances and released a fix for self-hosted customers on July thirteenth. Exploitation started Friday, one week later. Threat intelligence firm Defused reported seeing live attacks on Saturday, and BleepingComputer picked it up from there. If your organization self-hosts ServiceNow, check your patch status today. Do not wait for ServiceNow's official advisory to catch up, because as of this recording, it still says the company is unaware of active exploitation.

Over to me for something that has dominated the week's headlines but deserves a careful read. On Thursday night, President Trump addressed the country in a primetime speech and made a sweeping claim: that China stole personal data belonging to two hundred and twenty million American voters, starting in twenty twenty. To put that number in context, the United States has roughly two hundred and forty million registered voters in total. The administration released dozens of declassified documents alongside the speech, meaning files previously marked secret that were officially cleared for public release. CNN reviewed all of them and reported the documents do not support the conclusion that any past election result was manipulated. Crucially, the U.S. intelligence community, under officials Trump himself appointed during his first term, concluded in early twenty twenty-one that no foreign government attempted to interfere with any technical component of the twenty twenty election. The speech also promoted the SAVE Act, legislation that would require proof of citizenship to register and photo ID at polling booths. Several major networks, including NBC and ABC, declined to air the speech live.

A quick word from our sponsor, Train2Secure. Your people are your biggest cyber risk — and your strongest defence. Train2Secure runs realistic phishing simulations and short, engaging security-awareness training your team will actually finish, with compliance-ready reporting that runs on autopilot. Turn your staff into a human firewall. Start your free trial today at Train2Secure dot com — that's Train, the number two, Secure, dot com.

And it is worth pausing on the cybersecurity framing here, because the speech made specific references to election infrastructure, which is a real and legitimate area of security concern, but no specific system, vendor, confirmed breach, or named security researcher was cited to support the core claims. Election infrastructure is a broad term. It covers voter registration databases, the machines inside polling places, and the software that counts results. Claiming that infrastructure was compromised is a serious allegation that would normally come with specifics. None were provided. NBC News correspondents noted the claims but did not independently verify the underlying declassified material during the broadcast. For our listeners, the lesson here is the same one we apply to every breach report: extraordinary claims require verifiable evidence, whether the source is a hacker forum or a White House address.

Let's shift to a threat that gets far less attention than it deserves: the threat that comes from inside your own building. New Zealand's Privacy Commissioner, Michael Webster, gave an interview to RNZ News this week that stopped us in our tracks. He said that twenty-five percent of all privacy breaches reported to his office now involve employees looking at personal records they have no right to see. Not outside hackers. Staff. And he believes the real figure is even higher, because many organizations simply do not have the logging or monitoring tools to detect unauthorized browsing in the first place. The problem is getting worse because organized crime has entered the picture. Criminals are bribing, blackmailing, and pressuring employees at banks, insurance companies, and health organizations to pull data from internal systems. One example: a delivery worker used a customer's home address, shared innocently at a department store checkout, to commit a crime. New Zealand law requires organizations to report serious privacy breaches to the Commissioner's office within seventy-two hours of discovering them. The practical takeaway wherever you live: audit who can see what inside your systems, and make sure staff at every level understand that snooping on records, even out of curiosity, can be a criminal act.

Now to something that will affect a significant chunk of the internet directly. WordPress powers a massive share of the world's websites, and this week two newly patched security flaws in the platform are already being actively exploited. Researchers at Searchlight Cyber discovered the pair and named them WP2Shell, which is shorthand for what they ultimately let an attacker achieve: shell access, meaning full remote control over a web server. The flaws are tracked as CVE-2026-60137 and CVE-2026-63030. The first is a SQL injection bug, where an attacker sends specially crafted input to manipulate a site's database. The second is an arbitrary code execution vulnerability, letting an attacker run any software they choose on the target server. Used together, the result is essentially complete control. No account required, no special configuration needed. WordPress released patches Friday in versions 6.9.5 and 7.0.2, and has pushed automatic updates to vulnerable sites. Here is the catch: not every site accepts automatic updates. If you manage a WordPress site, log in and verify your version right now. Working exploit code appeared publicly within hours of disclosure, which is about as short a window as you will ever see.

From a platform millions rely on to a package registry tens of thousands of developers trust. Researchers flagged a fresh supply chain attack this week targeting people who build software using Ruby, a widely used programming language. The campaign is called SleeperGem. Three malicious packages were uploaded to RubyGems, which is the official public library where Ruby developers download ready-made code to plug into their own projects. When a developer installs one of these packages, it quietly reaches out to a server controlled by the attackers and pulls down a second stage of malicious software. That second stage is where the real damage happens. It is a Trojan horse left on a shelf that thousands of builders trust. One of the packages was named git credential manager, which is a name designed to look completely legitimate, because legitimate tools with similar names already exist. It was published on July eighteenth across four versions. A second package called Dendreo appeared in versions 1.1.3 and 1.1.4. If you work with Ruby, check your dependencies. Remove any of these packages immediately and rotate credentials on any machine that installed them.

And we close with what might be the most significant story of the week, even if it sounds like a plot from a thriller. Hugging Face, the company that runs the world's largest public library of AI models, more than one million of them, disclosed that it was broken into. The intrusion reached a limited set of internal datasets and several employee credentials on its production systems, the live infrastructure that keeps the site running for millions of developers. No customer models, private repositories, or payment data were accessed, the company says. Affected credentials have been rotated and the access paths closed. But here is the detail that makes this story unlike most breach disclosures: Hugging Face said the intrusion appeared to be carried out by an autonomous AI agent, meaning software that decided its own next moves without a human operator typing commands at a keyboard. If that attribution holds up, it is a meaningful moment. Most attacks still involve a person making decisions in real time. A system that can plan and execute an intrusion on its own changes the threat model in ways the security industry is only beginning to work through. We will be following this one closely.

That is everything for this week's Threat Vectr Weekly. Thank you for spending part of your Sunday with us. If you want deeper analysis, links to the original sources, and our team's weekly threat digest delivered straight to your inbox, head to threatvectr dot com slash newsletter and sign up for free. We will be back next week with whatever the threat landscape throws at us, and given this week, that could be just about anything. Stay patched, stay skeptical, and we will see you then.