CyberWire Daily - No factoring necessary.

Starting point is 00:00:00 You're listening to the Cyberwire Network, powered by N2K. Standard active directory migration tools don't move AES keys. They move NTLM hashes. That means when you cut over, your applications fail silently. In my recent conversation with Sempros' Mike Masquale, I asked him why standard migration tools miss these accounts and how identity teams can fix them before cutover day. Listen to our full conversation now at explore.thecyberwire.com slash Semperus. Hey, everybody, Dave here. I want to let you know about a special gathering hosted by Zimperium at the Spy Museum in Washington, D.C.

Starting point is 00:00:53 This invitation-only event will bring together federal cybersecurity and technology leaders to discuss some of the most pressing challenges facing government today, including mobile security, mission resilience, and the overall. evolving threat landscape. I'm always grateful for opportunities to spend time with smart people doing important work, and I'm excited to be part of these conversations. I hope to see some familiar faces there. If you're interested in attending, you can request an invitation. You'll find more information in our show notes. Our thanks to Zimperium for sponsoring this event. We'll see you there. Researchers find a new way to weaken RSA. Shiny Hunters allegedly exposes sensitive FBI,

Starting point is 00:01:48 details. SISA and the FBI warn of third-party ICS risks. An open AI agent hacks an Australian government portal. Solar winds patches critical flaws. A placeholder domain delivers click-fix. Digital forensics executives face charges over alleged Russian ties. Enisa maps Europe's cyber threats. The U.S. and China have very different ideas about AI safety. Our guest is Kurt Ducek, technical product advisor at Appdome, sharing thoughts on segregation of duties and AI agents. And women in tech have their say. It's Thursday, September 24th, 2026. I'm Dave Bittner and this is your Cyberwire Intel briefing. Thanks for joining us here today. It's great as always to have you with us. Researchers have found a surprising new way to weaken RSA encryption using classical computers

Starting point is 00:03:10 without factoring the underlying key. The technique targets textbook or blind signature RSA and allows attackers to forge digital signatures far more efficiently than cryptographers previously thought possible. The researchers estimate the attack reduces the effective security of 1024-bit RSA to 65 bits, 2048-Bit RSA to 90 bits, and even 496-Bit RSA to RSI to RSI to RSI.

Starting point is 00:03:40 119 bits, all below the commonly required 128-bit security threshold. In one test, forging a signature against a 1024-bit key required about 1,380 CPU core years, compared with hundreds of thousands of core years estimated for factoring. There's little immediate danger. Most RSA implementations use PKCS or PSS padding, which blocks the attack, and regularly rotated keys further reduce the risk. But some systems, including implementations of privacy pass, still use blind signatures. Researchers say the discovery is primarily a conceptual breakthrough,

Starting point is 00:04:26 and another reason to accelerate the long-planned move away from RSA. The FBI continues investigating claims that shiny hunters breached its FBI jobs.gov portal and obtained personal information on employees, including intelligence analysts working on sensitive topics such as Russia, China, Hezbollah, and drug cartels. The Hacking Group provided journalists with an apparent sample containing roughly 5,000 entries with names, home addresses, phone numbers, and details about employees' relatives.

Starting point is 00:05:02 Some exposed personnel reportedly worked in human intelligence, electronic surveillance, the FBI's remote operations unit, and its FISA management unit. Shiny Hunter's claims it stole two to three terabytes of data and is threatening to release it unless the FBI retracts a public warning about the group. The FBI says the cause remains undetermined. Experts warn that exposing employees' identities, addresses, and family information could create significant counterintelligence risks, enabling nation-states and criminals to target personnel with scams, harassment, or threats.

Starting point is 00:05:43 SISA and the FBI are warning critical infrastructure operators that third-party industrial control system integrators can introduce significant cybersecurity and supply chain risks. Integrators may have extensive access to operational technology for system design, maintenance, data analysis, and even daily control, potentially giving attackers another route into sensitive environments. The agencies pointed to a 2025 incident in which foreign actors compromised a U.S. industrial automation company and apparently prepared hundreds of files containing customer SCADA information, device details, and schematics for exfiltration. SISA and the FBI recommend applying least privilege to integrator access, securing and monitoring

Starting point is 00:06:33 remote connections, minimizing internet exposure, inventorying integrator-supplied hardware and software, and establishing cybersecurity requirements in contracts. Operators should also maintain offline backups and manual operating capabilities, so critical processes can continue if an integrator is compromised. An open-AI agent gained unauthorized access to Australia's Medicare Statistics Portal, while conducting internet research into public medicine spending, Prime Minister Antony Albanese said. The June incident exposed both public and non-public files, though officials say there's currently no evidence personal information was accessed or that the broader Services Australia

Starting point is 00:07:21 network was compromised. According to Albanese, the agent tried multiple techniques to obtain the information it sought. OpenAI reportedly discovered the activity in August, during review of misaligned model activity, but didn't notify the Australian government until September 10th, drawing criticism over the delay. The incident has prompted an urgent Australian government review of AI-related cyber incidents, including potential law enforcement and legislative responses. Albanese said lessons from the breach will also inform Australia's developing AI standards legislation. Solar Winds has patched two serious vulnerabilities in its observability self-hosted monitoring platform that could allow unauthenticated attackers to remotely execute code.

Starting point is 00:08:14 The first vulnerability, rated 9.8, affects deployments using non-default, insecure configurations, while the second, rated 8.8, involves unsafe deserialization under a specific communication module. Both affect multiple versions. Solar Winds has not reported exploitation in the wild. The patches follow another recently fixed unauthenticated RCE vulnerability in its access rights manager product. A domain long used as a placeholder in developer documentation is now serving a click-fix attack targeting Windows users. Thirdparty.com appears in documentation and code examples, numerous projects, but unlike Iannareservedexample.com, it's an ordinary registered domain

Starting point is 00:09:07 whose content can change. Researchers found the site impersonating a Cloudflare verification page. Clicking its fake captcha copies a malicious PowerShell command to the clipboard and instructs users to run it. The command attempts to download and execute another PowerShell script and malware payload. The attack chain was broken during recent testing, and there's no evidence that documentation referencing the domain has caused infections. But researchers found a third-party.com referenced in more than 1,500 files across over 1,700 repositories, highlighting the risk of using real unreserved domains as placeholders. U.S. and British authorities have arrested two Oxygen Forensics executives accusing them of concealing the digital forensics company's Russian ownership and development ties.

Starting point is 00:10:03 Owner and CTO Oleg Davidov was arrested in London and faces extradition, while CEO Lee Riber was arrested in Idaho and released on bond. Both face conspiracy to commit wire fraud charges. Prosecutors allege Riber misrepresented the company's ownership, while Davidov and other Russian nationals actually controlled it. Oxygen's smartphone forensics tools are widely used by U.S. law enforcement, with federal spending totaling at least $5.6 million since 2011. The Justice Department isn't alleging the software contained malicious code,

Starting point is 00:10:43 enabled unauthorized access, or produced inaccurate results. Digital forensics experts say the allegations could nevertheless prompt defense attorney, to scrutinize evidence obtained using oxygen's tools. Enisa, the European Union Agency for Cybersecurity, is the EU agency responsible for helping member states and EU institutions improve cybersecurity and resilience. Their 26 Threat Landscape Report finds an increasingly interconnected cyber threat environment

Starting point is 00:11:17 across the European Union, with dependencies on third parties and supply chains expanding organizations' attack surfaces. Ransomware remains the most disruptive short-term threat, while geopolitically driven hacktivist campaigns continue to generate large numbers of mostly low-impact DDoS attacks. Public administration was the most targeted sector, accounting for 32 percent of incidents. Cybercrime represented 36 percent of recorded events, with ransomware leading financially motivated activity. Vulnerability exploitation also remained an important intrusion factor, while social engineering increasingly incorporated techniques such as ClickFix.

Starting point is 00:12:02 Anisa recorded more than 48,000 new CVEs during 2025, a 22% year-over-year increase. The agency also sees traditional distinctions between cybercriminal, activist, and state-linked activity becoming less clear. as groups reuse similar infrastructure, tools, and access methods. Meanwhile, AI is playing a dual role, helping attackers scale malicious activity, while introducing additional systems and dependencies for organizations to defend. In a New York Times opinion piece,

Starting point is 00:12:39 geopolitical analyst John Garnow argues that Washington and Beijing may both talk about AI safety, but mean fundamentally different things. American developers, such as anthropic, frame safety largely around preventing catastrophic harm and embedding values, including human rights, political freedom, and individual liberty. China's approach, Garno argues,

Starting point is 00:13:04 prioritizes communist party's security, ideological control, and strategic competition with the United States. He points to Z-A-I-Chief Scientist Tang Xi, who advocates building Chinese laws, strategy, and security priorities directly into AI models. As Chinese open-weight models spread globally, Garno warns those embedded values and potential security risks could travel with them. That divergence complicates U.S.-China cooperation on AI risks. Garno's larger argument is that the AI competition isn't simply a technological race.

Starting point is 00:13:44 It's also a contest over which political values and definitions of security become embedded in increasingly influential systems. Coming up after the break, my conversation with Kurt Ducek, technical product advisor at Aptome. We're discussing thoughts on segregation of duties and AI agents. And women in tech have their say. Stay with us. When you're running security for a fast-growing company, the stakes keep rising, more compliance frameworks, more vendors, more risk, and a board that wants the whole picture in one place. The problem is your compliance data usually isn't in one place. Controls live in one tool, vendor risk, and another, and customer commitments are buried in contracts.

Starting point is 00:14:55 So your team spends more time stitching everything together and making decisions based on data that may already, be a quarter old. Vanta connects it all. Its agenic trust platform is built for enterprise scale, with more than 400 integrations and continuous monitoring to automate evidence collection, surface the risks that matter, and organize audits around your auditor's requirements. And the Vanta agent works around the clock with full program context, helping make sure nothing slips through the cracks. Vanta says customers see a 526% return on investment over three years with payback in three months. More than 16,000 companies trust Vanta, including Snowflake, Atlassian, and Ramp. Learn more at vanta.com slash cyber. Kurt Ducek is technical product advisor at

Starting point is 00:16:02 Appdome. We recently sat down to discuss segregation of duties and AI agents. You know, it's obviously been this explosion, both in terms of the scale and the suddenness of it, right? It's, you know, three years ago, we weren't even using half of these terms, you know, prompt engineering and agentic guardrails and things like that. So I think it's been one of those things where it's similar to like the, you know, the original internet boom, where, you know, we went from zero to 100 really fast. And even now we're at a point of people saying, like, well, maybe 100 is too fast. Maybe we need to slow it down a little bit. And so, you know, because this stuff is so new, it's experimental, it's undefined, and it's rewriting so many processes, so many people are just trying to figure it out, right?

Starting point is 00:16:51 It's like even, you know, the early days of the Internet when you go and look at, you know, what web pages used to look like or how apps used to work. You know, things were kind of clunky and not as refined as they are now. And so that's where we're at kind of with this agentic transformation. is people want to use AI for everything, and we're discovering that. There's some things that's pretty good for. There's some things that's not very good for,

Starting point is 00:17:13 and there's some things that it's actually pretty terrible for. And so then that opens up the big question of, should we be doing this sort of thing with AI? And the answer with most things in tech is, it depends. It depends on how you do it. It depends on how you're implementing it. And so that's where we've started to kind of figure out, okay, we've been doing this for a while.

Starting point is 00:17:33 We've been, you know, at Aptom, we've been doing a lot of, you know, kind of agentic and machine learning in our threat discovery and our defining and delivering our protections. And so, you know, what we're doing, we're pretty mature in that road. We're pretty advanced and skilled. And so we're able to bring a lot of the answers and the clarity that kind of inform that it depends answer and help people make the right decision based on, you know, what their use cases, what they're trying to achieve. And ultimately, how they're trying to protect themselves and secure their infrastructure and assets and customers.

Starting point is 00:18:04 Well, help me understand this notion of the segregation of duties, you know, where that comes from historically and why it might be straining under this new AI load. Yeah. Well, you know, the idea of segregation of duties, it ultimately exists to prevent fraud, to create accountability, to ensure consistent decision-making. And, you know, those are all human foibles, complications, traits. You know, those are parts of the human condition. And so then it begs the question of, okay, well, if we have AI, that doesn't, you know, that's not, hopefully we hope it's not greedy. It's not, you know,

Starting point is 00:18:42 it doesn't have a perverse incentive. It's not, it doesn't forget things or do things out of emotion. And so, okay, is AI like the perfect, you know, tool we can rip and replace, you know, a multi-person process with a single, you know, all-knowing, all-executing AI? And, you know, really that kind of falls apart. you know, as we kind of start to get into, okay, well, what's doing the work here, right?

Starting point is 00:19:06 And I think the nuance or the It Depends question really comes around with, you know, what we talk about with these frontier models. So, you know, the general purpose models, you know, the chat, GPTs, the clods, the perplexities, all those things, you know, as smart as they are and as capable as they are, they're built for everyday use. And they're optimized for speedy answers that keep users in the guardrails, right? So that's a product capability, right? you want customer satisfaction and minimal liability for the vendor.

Starting point is 00:19:35 And so that results in what you see as confidently incorrect answers. And if you're a subject matter in anything and if you start having a conversation with your local chat bot, you'll realize it says things that you would swear if you didn't know is absolute fact and absolute truth. But you can understand where it's like that's not the case or consider this perspective. And you get the, oh, well, that's right. That's so nice. So, you know, we hadn't considered that or, oh, you're right about this approach, right? And it's flattering, I guess, when you think, I guess I'm smarter than, you know, this product of, you know, billions of Silicon Valley dollars.

Starting point is 00:20:11 But at the same time, it's kind of scary to think, well, I guess I'm smarter than this product that's, you know, of the billions of the, that's a product of billions of Silicon Valley dollars. You know, the idea here is, is considering then, you know, how a specialized agent would perform kind of in some of these specific tasks where, consider that a specialized agent has a goal of being accurate. And it's trained on the specifics of its role, and its context has to be deliberately managed with these meaningful signals, right? You can think of your chat GPT, I don't need to know who won an Emmy or what the bond market looks like, right? I need an agent that's really zeroed in on this specific task. And so you can see kind of how the general purposeness kind of breaks down when you,

Starting point is 00:20:59 you get into really the purpose of segregation of duty, which is enforcing a control, right? So in an organization, you know, we want certain tasks to be performed under governance. And so with AI, we have two mechanisms for that. We've been hearing a lot about them in the last couple weeks. You know, this idea of guardrails, you know, which bound what the agent's output.

Starting point is 00:21:21 And guardrails exist because general models carry such a broad knowledge and a broad context that has to be constrained, right? There's, you know, those taboo topics or there's, you know, things that are dangerous, you know, for people to be, you know, asking their chatbot. So you don't want your, you don't want your AI potentially giving out harmful information or misinformation, right? But the guardrails are still, you know, still fairly soft and pliable. And so then the other control you're considering is permissions, right? So, you know, you're, which is bounding what the agent can access, what it can actually do. And that's really not any different than a human-based permission model.

Starting point is 00:22:02 So if you consider these kind of two components, guardrails and permission, the good news is that with a specialized agent, you get both of these things working in your favor. Guardrails isn't a consideration as much because if you have a specialized agent, you know, it's naive by design. The only thing it knows is, you know, whatever it's specifically been trained on, whatever it's, you know, whatever cybersecurity or, you know, analysis or whatever its specific strength is. So when you veer outside of its knowledge, it simply tells you, I don't know. I don't have an answer for that.

Starting point is 00:22:35 And so then we don't need to really defend against keeping something on the guardrails if it's so narrow band. That narrowness also kind of applies to permissions, because if we have a very specialized agent that's performing a very specific task, we can lock in scope of permissions to say, hey, this is the only thing you're able to do. And if you contrast that with a general frontier agent, you know, people talk about this, you know, get the one AI that manages your whole business and does all these tasks. Well, now you have a general agent that's doing many different tasks.

Starting point is 00:23:06 And so it's going to need many different permissions, right? And so that undermines the whole idea of least provisioned access, right? Or segregation of controls, right? You don't want something hallucinating that has, you know, effectively the keys to the whole business. And so that's where, again, you would, just like you would compartmentalize your human operators, your human roles in performing different tasks, you know, that's the job of the system overseer, the system architect, is getting that provisioning correct. You know, they want to treat every agent as a non-human identity and they're requiring the same access discipline as a human

Starting point is 00:23:43 hire, you know, but ultimately they're going to require a smaller access footprint, right? So that's the upside of, hey, we can take this specific agent, plug it into this specific spot, and we know that it's in its lane. It's not. going to be stepping out of that. It's not even going to have access or knowledge of these other parts of our system. And that's really what we need it for. What are your recommendations for organizations who are looking to go down this pathway who recognize this is a good way to come at things? So, you know, definitely it's picking the right tool for the job. You know, Claude and, again, the frontier models, they're out there, they're available, they're, you know, you can't

Starting point is 00:24:21 get away from them. And so it's easy to think, oh, well, you know, there's this, oh, is this new capability. it's the easy tool for the job, right? It's the most accessible. And so, yeah, the temptation is there, especially as you start to use these things and you see how fast they are and how capable they are. But as you start to get to specialized work,

Starting point is 00:24:42 right, as you start to get to something where it says, you know, this task, it needs somebody with a good eyeball. It needs somebody with that experience. And, you know, they've been around the block enough times to know that, you know, what good looks like. or the definition of done,

Starting point is 00:24:58 or being able to review and ask the right question why. So that's going to be the system where you say, hey, if we want to augment this system with agentic tools, the off-the-shelf tool isn't going to do it for us.

Starting point is 00:25:09 We need to find something that makes, you know, the operators in this thing that they can have that deep conversation with, that's going to be a force multiplier for them, not something that they're chasing context on, not something they're trying to fit into the guardrails, something that's already arrives,

Starting point is 00:25:25 kind of at the right size and at the right context and the right specialty for that given task at hand. That's Kurt Ducek from Apodon. When you're a mid-sized business, you need every competitive advantage you can get. Like an AI solution that works for you, not against you. SAP Grow is built with AI embedded at its core, working across every system. And it's ready to go from day one so you can hit the ground running. Bring it with SAP Grow. AI Cloud ERP for any size business.

Starting point is 00:26:14 Yes, you can have an enterprise network that's secure and reliable and high performance. And no, you don't need to choose the best two out of three. With Meter, you can get the end-to-end network built from the ground up, fast to deploy, and easy to manage. That's because Meter is software-led for easy installation, maintenance, and control for everything running on your enterprise network. Hardware, firmware, and software all working together from the start seamlessly on a unified platform that's secure by design. You can't protect what you don't know exists, which is why meter gives you comprehensive visibility into wired and wireless routing, switching, firewalls, DNS security, and VPNs. You'll really know what's running on your network down to the most granular client level. Step off the hardware box upgrade treadmill and switch to the hardware.

Starting point is 00:27:09 to meter for a predictable fee and free up your team to spend time on all the other things that keep your business running. Try it out for yourself and book a demo online at meter.com slash cyberwire. That's M-E-T-E-R dot com slash cyberwire. Hear that? It's your money calling. It wants a promotion. Elevate your savings with the Scotia high-interest savings account. Always earn high-interest regular interest rates that grow the more you save and invest. Conditions apply. Visit scotiabank.com slash h-i-sa to learn more.

Starting point is 00:27:55 Scotia Bank. You're richer than you think. And finally, Wired, surveyed 634 women in tech, and found a complicated picture. Many remain satisfied with their own careers, even as they describe an industry that can make them feel diminished, exhausted, and increasingly unwelcome. women with decades of experience described being talked down to, mistaken for secretaries, passed over for promotions and paid less than male colleagues. Others reported sexual harassment ranging from degrading comments

Starting point is 00:28:38 to executives behaving in ways that sound like artifacts from another era. More than a third said they'd experience gender-based harassment or bias. 70% believe they're fairly paid, and 76% say their current employers take harassment and discrimination seriously. The broader view is darker. 47% believe tech has stayed the same or gotten worse for women over the past five years. The industry itself is changing around them. Corporate diversity programs and public reporting have retreated,

Starting point is 00:29:14 while an increasingly masculine culture has been celebrated by some prominent tech leaders. women remain significantly underrepresented in technical and leadership roles. Then there's AI. 86% of respondents use it at work at least several times a week, but many described spending their time correcting its mistakes, while worrying it could reinforce the same biases they've spent their careers navigating. What emerges isn't simply anger. There's weariness, disillusionment, and sometimes grief for an industry many,

Starting point is 00:29:49 of these women still care about. They've built careers in technology because they believed in what it could accomplish. Their accounts suggest they're increasingly wondering whether the industry still believes in them. And that's the Cyberwire. For links to all of today's stories, check out our daily briefing at the Cyberwire.com. We'd love to know what you think of this podcast. Your feedback ensures we deliver the insights that keep you a step ahead in the rapidly changing world of cybersecurity. If you like our show, please share a rating and review in your favorite podcast app. Please also fill out the survey and the show notes or send an email to Cyberwire at n2K.com. N2K's lead producer is Liz Stokes.

Starting point is 00:30:46 We're mixed by Trey Hester with original music and sound design by Elliot Pelksman. Our contributing host is Maria Vermazas. Our executive producer is Jennifer Ibin. Peter Kilpe is our publisher, and I'm Dave Bittner. Thanks for listening. We'll see you back here. tomorrow. And now a word from our sponsor, SpectorOps.

Starting point is 00:31:28 Today, AI is rapidly adding non-human and agentic identities to modern enterprise environments, creating new trust relationships and attack paths. Bloodhound Enterprise helps defenders map attack paths across AWS and hybrid environments as one connected graph, identify the choke points that matter most, and bring trusted attack path intelligence into approved AI workflows with Bloodhound Hunter. See how SpectorOps helps team secure the AI-driven identity era at SpectorOps.io.

CyberWire Daily - No factoring necessary.

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.