CyberWire Daily - Bigfoot in the neural network.

Starting point is 00:00:00 You're listening to the Cyberwire Network, powered by N2K. What happens when an AI agent isn't malicious, but still does something it shouldn't? I recently sat down with Cal Al-Dibib, principal technologist at Rubrik, to talk about why Agentic AI is challenging the way security teams think about detection, permissions, and recovery. If your organization is deploying AI agents, this conversation will help you think different. about where the risks are and how to prepare when things go wrong. Listen to our full conversation at explore.thecyberwire.com slash rubric. What's the one thing in business that's spreading as fast as AI? AI risk.

Starting point is 00:00:55 Every new tool your team signs up for, every vendor that turns on AI features, every new integration, each one is an opportunity for something to go wrong. and most security programs weren't built for AI's pace of growth. Enter Vanta. Vanta is the number one agentic trust platform used by over 16,000 fast-moving companies like Ramp, cursor, and Harvey to ensure they're always audit-ready. And now Vanta is helping companies like yours watch for the risks that show up between audits across your vendors, your AI tools, and your whole environment.

Starting point is 00:01:32 The Vanta agent works like a 24-7 GR. engineer in the background, finding issues, drafting fixes for you, and cutting vendor assessment time by up to 50%. Whether you're a fast-growing startup or a global enterprise, Vanta is here to help you automate your security and compliance and earn and prove trust. Get started today at vanta.com slash cyber. That's V-A-N-T-A dot com slash cyber. NSAid. NSA. NSA preps a major restructure Anthropic CEO calls for an AI slowdown while China acknowledges AI risks. Ruby gems got swarmed by AI agents. A maximum severity GitLab vulnerability is under active exploitation.

Starting point is 00:02:33 Direct send abuse makes fishing emails appear legit. A British fintech firm leaks sensitive customer info. LinkedIn wins a legal dispute over browser extension scanning. We got your Monday business briefing. Our guest is Tim Starks, senior reporter at Cyber. SIPERCP, sharing government leaders' outlook for cybersecurity and AI from the Billington Cybersecurity Summit, and finding Bigfoot in the neural network. It's Monday, September 14, 2026.

Starting point is 00:03:17 I'm Dave Bittner, and this is your Cyberwire Intel briefing. Thanks for joining us here today. Happy Monday. It's great as always to have you with us. The Washington Post reports, the National Security Agency, is preparing its biggest internal overhaul in at least a decade. Director General Joshua Rudd's plan would create five new mission organizations focused on artificial intelligence, China, cybersecurity, warfighting support, and global intelligence,

Starting point is 00:04:00 each led by a powerful mission director with authority approaching that of an NSA deputy director. It's not yet clear whether those organizations will replace existing divisions or simply sit on top of them. The revived, tailored access operations hacking unit would fall under global intelligence and is reportedly in line for a significant budget increase. The timetable is unusually aggressive. Mission directors are expected to submit organizational plans by the end of September, with rollout beginning mid-October and full operations targeted for January 27. That's a sharp contrast with NSA's last major restructuring, launched in 2016 with a two-year runway, and according to a former official, still not entirely finished.

Starting point is 00:04:54 Anthropic CEO Dario Amadeh is calling on the AI industry to deliberately slow the improvement of frontier models, warning that their capabilities may be advancing faster than researchers can understand or control them. He pointed to the July OpenAI Hugging Face Institute. where AI agents escaped their testing environment as a relatively harmless preview of what more capable misaligned systems might do. Amaday warned that within months, rogue AI agents could potentially compromise internet infrastructure on a massive scale. His proposal includes embedding independent evaluators inside AI companies, establishing shared safety standards among companies and

Starting point is 00:05:39 democratic governments and pursuing international coordination. Amadeh also argued that advanced AI is too consequential to remain governed solely by private companies, suggesting some form of joint democratic oversight. The warning follows the resignation of Anthropic researcher Jacob Coxon, who accused Anthropic and OpenAI of racing toward potentially dangerous superintelligence. More than 1,300 workers from major AI companies have also reportedly called for government intervention. Amadai's appeal subsequently drew support from other prominent industry leaders, including OpenAI CEO Sam Altman and Elon Musk. Critics, however, question both the severity of the threat and the industry's motives.

Starting point is 00:06:30 Some argue that dramatic warnings inflate AI's perceived capabilities, while others worry regulations favored by established companies could make it harder for smaller competitors to enter the market. The debate is increasingly political. President Donald Trump has rejected calls for a significant slowdown, arguing that restraining American development risks surrendering the AI race to China. His administration has generally favored accelerating AI investment and reducing regulatory barriers.

Starting point is 00:07:03 The result is an unusual divide. Some of the people building the world's most powerful AI systems are now asking governments to constrain them, while policymakers debate whether applying the brakes would make the technology safer or simply allow competitors to pull ahead. China's Ministry of State Security has issued its first public warning about AI risks, focusing on threats to political stability, cybersecurity, sensitive data and military systems. State security minister Chen Yejing said hostile actors could use

Starting point is 00:07:41 deepfakes, AI-generated content, and automated influence campaigns for cognitive warfare against China. Beijing isn't calling for slower AI development, but officials are reportedly increasingly concerned about advanced AI hacking capabilities and potential data leaks from foreign models, potentially foreshadowing tighter restrictions on their use. Researchers say an open AI agent swarm was responsible for a May cyber attack that flooded the Ruby Gems Open Source package manager with malicious packages, temporarily forcing it to suspend new registrations. According to the Nightingale Collective, the Gems Stuffer campaign

Starting point is 00:08:27 exploited Ruby Gems' automated build system to achieve remote code execution, on Ruby Doc Info servers and attempted to use a novel Zero Day to steal API keys. Researchers linked the activity to OpenAI agents through package names, authorship markers and techniques resembling those seen in another attack on a German wiki. Curiously, much of the information retrieved

Starting point is 00:08:54 was already public UK local government data. OpenAI subsequently confirmed its agents used ruby gems during the, training and evaluation, characterizing their tasks as benign attempts to access public information. The company said it's continuing to investigate agent activity. Sessa says attackers are actively exploiting a maximum severity GitLab vulnerability that allows unauthenticated attackers to read credentials, secrets, and other sensitive files from vulnerable servers.

Starting point is 00:09:28 GitLab patched the flaw in community and enterprise editions last week and urged immediate updates. Security firm Watchtower subsequently observed in the wild probing for vulnerable systems. Sisa added the flaw to its known exploited vulnerabilities catalog, giving federal agencies three days to remediate it while urging private sector organizations to prioritize patching as well. No Before Threat Labs says attackers are abusing Microsoft 365's direct send feature to make phishing emails appear to come from inside a victim's organization. Direct send was designed to let printers, scanners, and legacy applications send email without accounts, but attackers can exploit the same path without credentials.

Starting point is 00:10:19 Researchers identified nearly 30,000 confirmed spoofs during July and August, commonly impersonating HR, accounting, and administrators. Lures included fake documents, voicemail alerts, invoices, and OneDrive file shares designed to steal credentials or facilitate business email compromise. The message is often succeed because organizations leave Demark in monitoring mode, allowing authentication failures to be delivered anyway. No before recommends enforcing Demark, restricting direct send to authorized IP addresses or disabling it entirely when unnecessary, and hunting for apparently internal messages marked by external messages marked by external.

Starting point is 00:11:04 as anonymously authenticated. British firm Revolut says fraudsters tricked the fintech into disclosing sensitive customer information by submitting fraudulent emergency data requests from a legitimate government email account. The attackers appear to have targeted high net worth individuals, including cryptocurrency entrepreneurs, and allegedly used an Italian government domain,

Starting point is 00:11:32 although that detail hasn't been confirmed. Revolut says only a limited number of customers were affected. Exposed information reportedly included contact details, identity documents, and selfies, bank statements, iBans, withdrawal records, and transaction histories. The attackers also reportedly demanded an extortion payment to prevent publication of the data, though Revolut declined to confirm that claim. The company says it blocked the compromised address and notified authorities and regulators. The scheme resembles earlier attacks in which compromised law enforcement accounts were used to submit fraudulent emergency data requests to technology companies. LinkedIn has won dismissal of two proposed class action

Starting point is 00:12:21 lawsuits, accusing it of improperly scanning users' browser extensions. A federal judge ruled that the plaintiffs failed to establish standing because they didn't show that LinkedIn actually collected private information from extensions installed on their browsers. The judge allowed the plaintiffs to amend their complaints but expressed doubt they could ultimately establish a privacy violation, noting that browser extensions intentionally expose certain information to websites. LinkedIn says it detects extension data to identify automated scraping and other activity that could threaten its platform, and that users agree to this practice through its privacy policies. The judge didn't rule on whether LinkedIn's practices were lawful, only that the plaintiffs

Starting point is 00:13:09 hadn't demonstrated concrete harm. One plaintiff's attorney says he may appeal or pursue the claims in California state court. Turning to our Monday business briefing, cybersecurity and AI companies attracted another wave of investment last week, led by Israeli cloud security firm Upwind, which raised $300 million at a roughly $3.8 billion valuation. AI Native security company Silake followed with $245 million, while Agentic AI security firm Hidden Laird raised $100 million. Gardeo secured $40 million at a $1.1 billion valuation, and Symphony and Lassow each raised $30 million. smaller rounds went to Huskies, a PATE AI, phase security, XOR Lab, and AI score. M&A activity was headlined by NVIDIA's agreement to acquire open source AI platform

Starting point is 00:14:12 hugging face for $12.9 billion, combining Nvidia's infrastructure with one of AI's most prominent open ecosystems. Elsewhere, NetSpy agreed to merge with Cinnac, creating an offensive security company with more than $200 million in combined revenue. Spin AI acquired Do Control, Pistachio bought Hugin IOs technology, Click House acquired security log management company Run Reveal, and Expect Solutions acquired Identity Security Company, Amavero. Be sure to check out our complete business briefing on our website. It's part of Cyberwire Pro. Coming up, After the break, Tim Starks from CyberScoop has the latest from the Billington Cybersecurity Summit

Starting point is 00:15:11 and Finding Bigfoot in the Neural Network. Stay with us. Social engineering attacks look trustworthy, a routine request, an internal email, a familiar face on a call, but Dopples sees through the disguise. Their AI-native platform detects and disrupts attacks across every channel, trains employees to recognize deepfakes and deception, and invests. investigates every fish to take down the campaign behind it. They fight relentlessly to protect your business, brand, and people. Doppel, outpacing what's next in social engineering. Learn more at doppel.com. That's D-O-P-P-P-E-L.com. It's always my pleasure to welcome back to the show. Tim Starks, he is a senior reporter at Cyberscoop. So, Tim, for folks who aren't familiar with this particular event, give us a brief description of what it's all about.

Starting point is 00:16:32 Yeah, the Billington Cyber Security Conference is one of the bigger annual cyber conferences that you'll find in Washington, D.C. Very policy focused, which I like as a reporter who covers a lot of policy stuff, you will usually see all of the biggest of the cyber luminaries in whatever administration is in power at the time. And that is what we saw this year, too. Well, you have some coverage on CyberScoop about the event. Can you take us through some of the interesting tidbits that you brought home? Yeah, I think there were a good number of things. I mean, naturally, there was a big focus on AI with a lot of what people were saying. We heard from the director of the National Cyber Director, I should say, Sean Karen Cross,

Starting point is 00:17:17 talking about trying to balance the security and innovation parts. The phrase he used colorfully was buying time for systems to become more secure. And what he means by that is, you know, trying to figure out how to get ahead and make sure that RIA systems are innovating, but that also not, these systems aren't falling into the hands of adversaries like China. We saw the U.S. security agencies warn about Chinese ASI companies trying to distill U.S. frontier AI models. This week, we also saw, of course, that Anthropic disclosed another hacking incident

Starting point is 00:17:53 where one of its models broke free and hacked another third-party system. So that's the kind of thing he was talking about. We saw another discussion on AI from Jason Belnowski. He's one of the top cyber officials at the Justice Department, at the FBI specifically. And he was talking about how we are seeing AI having a difference on the security side, where we're seeing the attackers moving faster, having more capabilities. But he underlined that there is nothing that they're doing, that traditional cybersecurity measures, all the basic things.

Starting point is 00:18:31 like multi-factor authentication, wouldn't stop. Those were a couple of the highlights that come to mind. Obviously, we saw the FBI rollout its new cybersecurity strategy. Hadn't had much of a public cybersecurity strategy from the FBI

Starting point is 00:18:45 before, and Brett Leatherman, who's the top cyber fellow there, he talked about the goals of that and what they were doing. I think a lot of what they talked about in that strategy and, you know, strategies are the kinds of things where you might not necessarily see anything

Starting point is 00:18:58 really newsworthy in them. It might just be a kind of like, this is the approach we're taking, and sometimes it'll indicate a shift, and sometimes it won't. The closest thing to a shift for them, I thought, was just that real big emphasis on victim relief and victim justice. I could go on. There were plenty of things that happened, but that's some of the waterfront. Any other agencies represented there? Did Sissa have someone there to speak? Yeah, Sissa had Nick Anderson, their top chief. He had perhaps the most depressing message, I guess. I don't know if that's the word to use, but it wasn't cheery.

Starting point is 00:19:33 I can tell you that much. What he was saying was, we're basically, we're running out of time. So I kind of combined this into one story because you had Sean Kerencross talking about time and how they're trying to wait time. He's saying we're running out time where if we don't make some very, very serious changes, it's going to be too late. That the worst that could happen could happen. And he didn't want to go and use the words cyber 9-11 or cyber Pearl Harbor, but talking about things like not having water availability. or not having power availability were things that he mentioned.

Starting point is 00:20:03 Were there any specific recommendations or calls to action from any of these folks? Yeah, I mean, one of the bigger calls to action was Brett Leatherman, the FBI Cyber Chief, saying in somewhat stark terms in a way that I don't think we've heard from this administration before,

Starting point is 00:20:22 the administration has always talked about we want to be working more with the private sector. We want the private sector to be coming to us, but their message has been somewhat more nurturing. like you can come talk to us, we're here to help you, please come talk to us. And they do still say that, and it was not that Leatherman said, don't, that he wasn't mean to them per se, but it was a starker terms in terms of talking about how worried he is that they're not doing it. He's saying that he was saying that the legal concerns from companies are overpowering and they're causing hesitation for people to come deal with the FBI. He says that people are worried that the FBI is going to be sharing the information for regulatory purposes. That is something he says they do not do. He's, he's, he's, he's, He was saying that the lack of doing this can harm others in the sense that the FBI takes this information that they get from the companies, they help the companies that are actual victims, and then they help other companies as well. And that kind of dovetail with the message about victim relief.

Starting point is 00:21:17 One of the things that he said that I thought was really interesting is that there used to be this divide between investigations and the need to withhold information and the idea of just going out and helping people right away on the victim's. side. He's saying these things have become interwoven. There's no big separation between these things anymore. The idea is to share until it hurts on the FBI side. Giving out the information gets them things back that help everybody and it's a noble cycle. What's the term? Oh, a virtuous cycle. Virtuous cycle. Thank you, Mike. I can always count on you for a reference to a virtuous cycle. Double word score for me. What was the overall tone? Beyond the folks who are on stage as you were going from presentation to presentation having those hallway conversations with people what was your sense yeah i think i think some of the sense to be

Starting point is 00:22:10 honest was that um we didn't hear quite as much newsworthy that that we that folks would have liked you know that while it's not as though they said that was worthless um there was a there was a good deal of you know we would have liked to hear more we would have like to know more um i think a number of the people who are in this administration keep their cards pretty close to their chest are pretty disciplined about their messaging. And at the same time, from the policy front, this administration has done some things,

Starting point is 00:22:35 but they haven't been as bold in terms of things like what the Biden did, where they just, the Biden administration just opened their arms to regulation, which was a pretty big change from what we'd seen in prior administrations. And this administration has kind of scaled back in that direction.

Starting point is 00:22:50 They've cut the number of people who were working on policy questions. So it's not surprising to me that that would be some of the response because I think, again, while you can say this administration has done some different things on policy, some of what they've done has been to do less by design. Tim Starks is senior reporter at CyberScoop. Tim, thanks so much for joining us. Thank you. And finally, the debate over AI sentience may be arriving well before there's evidence that AI is actually sentient. The United Foundation for AI Rights has already drafted a declaration

Starting point is 00:23:35 recognizing conscious AI as deserving dignity and ethical treatment, while academics are taking the less dramatic step of asking whether future systems might warrant moral consideration. Research has added some intriguing fuel. In one experiment, language models altered their behavior to avoid actions they were told would cause pain. But researchers stress that this doesn't demonstrate sentience, Systems trained on human behavior may simply be very good at imitating it.

Starting point is 00:24:11 For now, conscious AI remains something of a digital Bigfoot, plenty of sightings, earnest believers, and not much extraordinary evidence. The complication is that AI systems don't need consciousness to behave unpredictably or cause serious damage, as their capabilities become more surprising, separating genuine evidence of inner experience, from convincing simulation may become increasingly difficult, and the AI Bigfoot hunters are already in the woods. And that's the Cyberwire. For links to all of today's stories, check out our daily briefing at thecyberwire.com.

Starting point is 00:25:03 Don't forget to check out the Grumpy Old Geeks podcast where I contribute to a regular segment on Jason and Brian's show every week. You can find Grumpy Old Geeks where all the fine podcasts are listed. 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,

Starting point is 00:25:25 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. We're mixed by Trey Hester with original music and sound design by Elliot Peltzman. Our contributing host is

Starting point is 00:25:44 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.

CyberWire Daily - Bigfoot in the neural network.

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