AI Will Save Us, or Not? – PSW #945

Paul Asadoorian

1. DIY Router On X86 E-Waste: OpenWrt And OPNsense

   I'm not sure I would use an old PC for this because:

   • There are too many things that can break (power supply, fans, old disks, etc...)
   • It takes up a lot of room
   • They typically only have one ethernet adapter

   I did some research to see what AI would come up with, it did an okay job. Specifically I would recommend a fanless mini-pc instead, however, there are so many with so many different specs I'd need to spend more time weeding through the options, but here are some highlights:

   • Protectli, Topton N150 and Qotom are solid options, but once you build them with the right specs they are sitting between $300 and $500 easily (more expensive due to increased RAM and storage prices)
   • Intel N100 is nice, a bit slower, and not as great at hardware support
   • I did specify that it had to run either OpenWRT or OPNSense and have hardware compatibility; there's a lot more to dig into here, and a Wifi dongle would be an extra cost

   What I thought was most interesting: You can buy the smallest cloud gateway from Ubiquiti for $129, and it will push Gig Internet: https://store.ui.com/us/en/products/ucg-ultra - If you add an AP, the cost is around $247. We're not supported or sponsored by Ubiquiti, but it holds for my recommendation, however, my only concern is how long the product will be supported before it needs to be replaced (software updates, patches, etc...). That being said, if you are on a budget, getting a free e-waste PC and sprucing it up to run as your firewall is a good option, but again only if you are on a tight budget.

2. InfraTrust Pulse September 2026: What to Patch First

   I did an entire webinar on this month's data, built from advisories issued recently for infrastructure devices, with some interesting insights to discuss:

   • Sometimes the vendor states the vulnerability has been observed being exploited in the wild, even before it lands on a KEV list
   • Attackers are going after the network edge management platforms, more bang for the buck
   • There are now two CERT/CC VINCE records for UEFI shells, one disclosed by my co-worker Stas, and one new one that is weird and references our Bombshell disclosure with Framework (it even lists a vulnerable shell checksum that matches exactly to the ones I disclosed to Framework). Framework fixed all of these a long time ago. This is now getting extra coverage...
3. D-Link DIR-822A Stack Overflow in udhcpd TR-111 Option 125 Parsing

   Summary: Source analysis found a stack-based buffer overflow in the D-Link DIR-822A's udhcpd TR-111 parser. The router uses strcpy() to copy attacker-controlled DHCP option 125 (Vendor-Specific Information) suboption data into 256-byte stack buffers, even though the data is length-delimited and may not be null-terminated. An unauthenticated attacker on the LAN can send a crafted DHCP message to crash the service, corrupt memory, and potentially achieve code execution, depending on the firmware build. The exact affected public firmware version still needs to be confirmed.

   Paul's take: This is the kind of bug that makes me wonder how many routers are still parsing weird enterprise DHCP features nobody remembers enabling. A DHCP server should absolutely know the length of the data it is copying, and using strcpy() on network input is a classic way to turn a harmless option into a stack smash. The practical impact is local network access, but that is not a very high bar on a home or small-office router. Patch it if a fix exists, no wait, nevermind, get rid of this router and see my story #1.

4. AI Hallucination Nearly Triggered a US-China Military Confrontation

   Summary: An AI-generated intelligence report reportedly falsely identified nuclear weapons components aboard a Chinese ship during the Iran war, nearly triggering a U.S. military interception. The report came from two AI-assisted steps with no meaningful verification in between, and the article says military aircraft were already airborne before the intelligence was challenged. The incident highlights the danger of confident AI errors entering targeting and military decision-making systems faster than humans can validate them.

   Paul's take: This is the part where everyone says, "The computer is wrong," while the computer is already moving military aircraft. It is basically WarGames, except the lesson from the movie was supposed to be that you do not let a machine turn a guess into a launch decision. AI does not need to become self-aware to start a war. It just needs to produce a confident wrong answer, format it like intelligence, and reach someone who is rewarded for moving quickly. The human in the loop needs to be able to challenge the output, not just click approve on a machine-generated escalation. Did we forget the lesson? Shall we play a game? Also, backdoors are not secrets...

5. Hackers Stole Flock’s Camera Software, Revealing How the Company Trac…

   Summary: Hackers removed a Flock Safety camera, copied much of its software and data, and recovered an encryption key stored on the device that unlocked videos and images of vehicle detections. The analysis found that the camera detects people, vehicles, license plates, bicycles, and other objects, generating roughly 1.6 million images from about 50,000 vehicles during the recovered periods. The research also showed how Flock's camera network feeds a broad searchable system used by thousands of agencies, while the cameras themselves showed storage failures, crashes, and reboot problems.

   Paul's take: This is the IoT version of leaving the keys inside the surveillance camera. Flock says the data is protected by on-device encryption, but if the key is sitting on the device, physical access changes the whole security story. The bigger concern isn't just one camera or one stolen image; it is the national network built from these local systems and the number of agencies that can search it. Also, the camera explicitly detects people, while the public conversation usually frames it as license-plate tracking. This is just confirmation that Flock is not acting in anyone's best interest, except their own, and certainly not even trying to protect anyone's privacy.

6. NVIDIA Infrastructure Controller Hit by 14 Flaws Enabling Code Execution and Privilege Escalation

   Summary: NVIDIA released Infrastructure Controller version 2.0 to fix 14 vulnerabilities affecting versions 0 through 1.9. The set includes a critical hard-coded credential flaw rated 9.8, SQL injection, missing and improper authentication, OS command injection, certificate validation failures, file path control, XML injection, denial-of-service issues, and exposed debug information. Depending on the flaw, attackers could execute code, escalate privileges, alter data, disrupt services, or access sensitive information.

   Paul's take: The management plane continues to be the place where vendors hide the worst surprises. A hard-coded credential with a 9.8 score in infrastructure management software isn't a minor bug, especially when the software controls high-value systems. Update to version 2.0, restrict access to the management interface, and isolate it from untrusted networks. The key point is that one management product can include several ways to lose control of the environment.

   A little background on NVIDIA Infrastructure Controller:

   • What it is: A control-plane application for managing NVIDIA data-center infrastructure and related services.
   • Who runs it: Organizations operating NVIDIA enterprise infrastructure, such as AI clusters, GPU servers, or associated data-center systems. Infrastructure and platform administrators typically run it, not ordinary GPU users or gamers.
   • Why they run it: To configure, provision, monitor, and manage NVIDIA infrastructure from a central system.
7. Should you care about an “AI slowdown?”

   Summary: Talos argues that an AI slowdown would not substantially reduce cybersecurity risk because current models are already capable of finding vulnerabilities and accelerating attacks. The bigger issue is that many organizations still lack basic security controls such as accurate asset inventories, identity management, least privilege, and network segmentation. AI can improve defense, but it cannot compensate for weak fundamentals.

   Paul's take: I agree with this one. The AI arms race gets all the attention, but attackers are still walking through exposed systems, stolen credentials, bad configurations, and unpatched infrastructure. You don't need the next model release to cause problems when current models are already good enough, and the target environment is held together with tape. Do the basics, then add AI.

8. Opus 4.8 At Home? – Running Local LLMs for IoT Hackers

   Summary: Matt Brown tests two NVIDIA DGX Spark devices as a local AI setup for IoT security research, reverse engineering, and firmware analysis. Using DeepSeek, GLM, and Qwen models with the OMP agent harness, he runs real workflows involving firmware extraction, Binary Ninja analysis, APK decompilation, Bluetooth packet analysis, and cryptographic flaw discovery. DeepSeek V4 Flash solved an Android and Bluetooth challenge locally in about 40 minutes, compared with 22 minutes through OpenRouter and 15 minutes with Opus 4.8.

   Paul's take: My tip: Even if you have underpowered hardware, run some local models on it and give them tasks; have them do the boring, dirty work, then take those results and finish your task(s) on a frontier model (even a lighter one). It's been working well for me. This is a great local AI setup that feels genuinely useful for security research, not just a chatbot running on a desk. It is slower than Opus 4.8, expensive, power-hungry, and still has a smaller context window, but the intelligence is close enough to run real firmware and reverse-engineering workflows without sending everything to a hosted provider. For me, the privacy and data-sovereignty angle matters most. Give this hardware another generation or two and local models become a serious option for people doing sensitive security research.

9. Quantum Computers Are Not a Threat to 128-bit Symmetric Keys

   Summary: Filippo Valsorda argues that quantum computers do not require organizations to replace AES-128, SHA-256, or other 128-bit symmetric cryptography with 256-bit versions. Shor's algorithm creates an urgent threat to public-key systems such as RSA, ECDSA, and ECDH, but Grover's algorithm does not make brute-forcing AES-128 practical because the attack requires enormous serial quantum computation and does not parallelize efficiently. NIST and Germany's BSI continue to consider AES-128 acceptable for post-quantum security.

   Paul's take: (Sam will have a better take) This is an important correction to the idea that quantum computers automatically cut every key in half. We still have plenty of real post-quantum work to do, especially with public-key cryptography, so wasting time doubling symmetric key sizes would create churn without improving security.

10. How One Twitch Chat Message Became Code Execution on a Streamer’s PC – SCRT Team Blog
11. 10 Lessons Reshaping Security After Black Hat and DEF CON 2026
    • I don't see AI making attacks or attackers more sophisticated, I see it making attacks more ubiquitous and faster, but sophistication has not improved much. For example, on Linux systems attackers are still using 30-year-old tactics, moving much faster, but not more sophisticated. For me, AI is providing speed.
    • AI agents are just another user identity, we talked about this last week
    • Prompt injection will continue, but don't forget there are so many more layers to AI infrastructure, and each one has its own attack surface
    • "A severity score does not show whether a weakness is reachable or what an attacker could access through it. Attack-path analysis connects vulnerabilities with identities, configurations, network reachability, and critical assets." - I think the how is the important part with this one, many startups have tried, many have failed, and AI is not the best at this right now.

    And this seems to be a re-occuring theme this week: "...speakers repeatedly returned to asset inventory, identity management, least privilege, configuration hygiene, testing and attack-surface reduction. AI makes neglecting them more costly. A forgotten server is more exposed when discovery is cheaper. An excessive permission creates more risk when an agent can use it at machine speed. Poor data governance becomes harder to contain when an autonomous system can search repositories and send information through connected tools. Returning to the fundamentals does not mean ignoring AI. It means applying established security principles to agents and strengthening the controls on which AI-assisted defense depends."

12. Nightmare Eclipse Drops New Microsoft Defender Exploit After Revealing Identity
13. pocs/cups/cups2root at main · v12-security/pocs
14. cups2root: New CUPS Zero-Day Escalates lpadmin to Root – Cyber Kendra
15. My Ethical Hacker EDC: Every Gadget I Actually Carry – Everyday Carry Blog – EDC Blog

    This post got me to get my Hak.5 pager up and running, which I did, but I haven't dug into it. There are so many gadgets, and so little time to explore them all. The ESP32 platform just keeps getting better and more interesting and continues to be the go-to device for those who want to tinker and build custom or even off-the-shelf hacking devices. I think Linux handhelds will become a thing, and M5 has the CardputerZero launching this November: https://www.kickstarter.com/projects/m5stack/cardputerzero/description - And I am excited for this and the FlipperOne.

Lee Neely

1. Notice and Apology Regarding a Data Breach Resulting from Unauthorized Access to Gyazo

   Summary: Helpfeel Inc. has confirmed that Gyazo, our image-sharing service, was subject to unauthorized access by a third party, resulting in the unauthorized disclosure of user information and certain metadata associated with uploaded images. We have blocked all access routes used in the incident and have completed remediation of the vulnerability that was exploited.

   Lee's Take: The service uploads screenshots and screen-recordings, providing users a sharable link for posting in social media, email/etc. The vulnerabilitiy was exploited on September 11th, and the activity detected later that evening with all access routes closed early on September 12th. Even so, the meta data of about 490 million images along with 23.62 user records had been exfiltrated. That is a lot of data in a short interval. Make sure you're considering just how quickly data can be exfiltrated from a compromised system, both when evaluating response timelines and setting up monitoring for unusual activity. Beyond changing passwords, this would be a good time for Helpfeel to roll out MFA for everyone.

2. International security agencies warn about North Korean hackers exploiting job seekers to steal crypto, data

   Summary: Security agencies in Australia, Germany, Japan, and the US have published a joint advisory warning of malicious activity conducted by a threat actor group with ties to North Korea. According to the advisory, the threat actor group, dubbed WaterPlum or Contagious Interview, "conducts cyberattacks by infiltrating unsuspecting job seekers’ computer networks, harvesting sensitive information, and stealing cryptocurrency. WaterPlum is victimizing individual IT professionals in Japan, the United States, Europe, and other countries." By posing as employers offering job interviews for software developers and other professionals, the group has reportedly infected more than 30,000 devices in more than 100 countries and stolen funds and cryptocurrency wallet credentials. They generally contact targeted individuals via social media and online job platforms, coax them to be interviewed or receive a call about potential work, and use a variety of social engineering tactics to trick the targets into running malicious code on their devices. The advisory includes the threat actors' known tactics, techniques, and procedures (TTPs), and notes in closing that "The techniques described in this advisory are only examples; actors continuously evolve and refine their methods [and advises individuals to] stay informed by monitoring alerts from domestic and international security agencies and by reviewing reports published by security vendors."

   Lee's Take: In other words, fake job interviews are the new social engineering lure. With the change in the job market, some of which is related to AI adoption displacing employees, I'm hearing more folks who are frustrated, desperately looking for any sort of response to their applicaiton. Unfortunately this isn't the callback you're looking for.  Beware of an interview asking to execute code, or download applications to your sysetms, to include copying/pasting of commands. The 9-page report from IC3 below includes mitigations for employers and job seekers, and includes useful references for more information.

   IC3 Guidance: https://www.ic3.gov/CSA/2026/260918.pdf

3. New CISA-hosted VINCE-NT system revealed for coordinated vulnerability disclosure

   Summary: The U.S. Cybersecurity and Infrastructure Security Agency (CISA) announced the launch of Vulnerability Information and Coordination Environment – New Technology (VINCE-NT) on Thursday, an upgraded, fully CISA-hosted system that will replace the original VINCE for coordinated vulnerability disclosure (CVD).VINCE has been used by researchers, suppliers and stakeholders since 2020 to coordinate the disclosure of vulnerabilities.

   Lee's Take: The prior VINCE sysetm was rolled out in 2020, and used for coordinated disclosure of vulnerabilities, sponsored by CISA and hosted by CERT/CC (CMU). The new system is hosted, sponsored, and managed by CISA.  With an improved UI, more standardized information and reporting, this should improve the advisory publication workflow. New capabilities also include secure file sharing and seamless collaboration capabilities and better automation, this should provide a needed shot-in-the-arm aiding both defenders and researchers.

4. Cyberattack hits University of Munich, potentially exposing student financial data

   Summary: Ludwig Maximilian University (LMU) in Munich, Germany, is recovering from a cybersecurity incident that compromised sensitive student data. Over the weekend, LMU said that the intruders accessed an IT system that contains enrollment information; the university is assuming that it "must assume that this data were in fact retrieved." The compromised information may include names, dates of birth, contact details, LMU email addresses, bank account data, and details about students’ courses of study and previous educational qualifications.

   Lee's Take: Universities are a target for ransomware and data exfiltration attacks. LMU has disconnected more than just the compromised IT systems because of possible trust relationships, resulting in greater impacts to the students. While impacting more users is not a popular move, making sure you have contained the incident is critical, which also means you need to have system interdependency information to properly build your incident response plan. Yeah, that can be hard to get, and the alternative, during an attack, can be worse.

5. LinkedIn wins court order blocking mass scraping of user data

   Summary: Sarah Wight, VP of Legal for Litigation, Competition, and Enforcement at LinkedIn, has posted that a federal court in California "has entered a consent judgment in LinkedIn's favor," prohibiting ProAPIs, Netswift, and their founder and CTO Rehmat Alam from scraping, selling, and transferring data from LinkedIn member profiles. The companies created millions of phony LinkedIn accounts through which they conducted their activity. The October 2025 complained alleged breach of contract; fraud and deceit; breach of the Computer Fraud and Abuse Act; breach of the California Comprehensive Data Access and Fraud Act; unlawful, unfair or fraudulent business practices; violation of the Lanham Act; misappropriation; and trespass to chattels.

   Lee's Take: The accounts were used to create a continuously updated database of all LinkedIn users, including member, school and employer, comment and posts which they sold to others, without consent or permission. LinkedIn did detect and delete the phony accounts created to harvest member account data within hours, but that was still sufficient time for information to be harvested. The fake accounts were created by the hundreds, if not thousands, daily, overwhelming the countermeasures, kind of a DoS attack.  This behavior is not unique to LinkedIn, who are actively working to detect and block this behavior much more quickly. With this ruling, I would expect other social media providers to go after similar activity on their platforms. While tempting to pull back on your social media presence, many employers are now using it as part of their selection process, and if you don't have a profile, which re-enforces your application, you're going to be passed over, instead you may want to dial back on secondary or shadow profiles.

6. SolarWinds Patches ARM Hard-Coded Key Flaw Enabling Unauthenticated RCE

   Summary: SolarWinds has released security updates to address a high-severity flaw in Access Rights Manager (ARM) that, if successfully exploited, could lead to an unauthenticated remote code execution vulnerability. The vulnerability, tracked as CVE-2026-28326, is rated 8.8 out of 10.0 on the CVSS scoring system. The issue affects all versions of Access Rights Manager 2026.2 and prior.

   Lee's Take: If I could bill someone a nickel for every hardcoded password, you know where I'm going. Bottom line, update SolarWinds ARM to 2026.2.1. Version 2023.2.4 and below were all EOL in May, so if you're on one of these, you need to jump. Read the Installation and Upgrade guide to identify all the steps you need to take to get to 2026.2.1 regardless of your starting point. If you're using SolarWinds ARM, you need to make sure it's on a separate server from your other SolarWinds platform products which is neither public nor Interntet facing.

7. Fake LastPass Authenticator GitHub repos push new Rapuncel infostealer

   Summary: Threat actors have impersonated more than 40 organizations to spread infostealer malware known as Rapuncel. The campaign was detected "on August 13, 2026, [when] the LastPass Threat Intelligence, Mitigation, and Escalation (TIME) Team identified a fraudulent GitHub organization impersonating LastPass Authenticator which redirected visitors to attacker-controlled infrastructure which ultimately served the installer." The attackers used search engine optimization (SEO) to elevate a GitHub page serving the fraudulent LastPass installer; the threat actors appear to have done the same thing for a fraudulent macOS LastPass application. LastPass writes that "the campaign delivered a Microsoft-attested kernel driver with zero VirusTotal detections that terminated 145 antivirus and EDR processes from kernel mode, enabling potential credential theft across browsers, cryptocurrency wallets, Discord, Steam, Telegram and Windows Credential Manager."

   Lee's Take: The attack involves a user attempting to download a package update, which instead, triggers a series of redirects, ultimately leading to the attackers' servers which provide a 148mb Zip file, which evades most security scans. Inside is an installer with a signed copy of the MS Visual Studio debugger which is renamed and configured to sideload the malicious vsdbg.dll DLL, then deploy the Rapuncel infostealer and Alinubx.sys kernel driver, which kills the AV/EDR software, while registering itself as the NVIDIA NvFSFIler service. To bypass app encryption in Chromium based browsers. the Infostealer injects a helper DLL and invokes its own Elevation Service. Rapuncel persists across reboots, so any EDR/AV products are killed again if reactivated. There are only a couple of things we can do here, first grab the IoC's from the LastPass blog, unt for them and block the listed domains, next remind users to only download from known good sources, and lastly, for systems compromised, users are going to need to update their credentials. ALL of them - from a known clean machine - any credentials stored in browsers; credentials for Discord, Steam, Telegram accounts, and anything in the Windows Credential manager. If they had a crypto wallet on the system, change those credentials too. Maybe a good time to have them enable any MFA options, including passkeys; they had been postponing. LastPass Blog: https://blog.lastpass.com/posts/lastpass-delphos-report-rapuncel-infostealer

8. Colorado Water Utilities Hit by Cyberattacks Targeting OT Systems

   Summary: Colorado Governor Jared Polis has disclosed that operational technology (OT) systems at two small water utilities suffered breaches in late August 2026, several weeks after similar incidents at other water utilities around the US made headlines. Both targeted utilities are private, and both serve fewer than 200 people. A spokesperson for the governor said the "incidents consisted of individuals changing equipment settings, disabling remote access and alarms, and altering pumping cycles." Both utilities address the security issues themselves; the incidents did not affect the utilities' ability to provide services, nor did they threaten public safety.

   Lee's Take: Thus far the recent attacks on water utilities have not disrupted service to customers. But, if they are messing with the control systems, testing the compromised capabilities, it very well can happen. Consider this a mulligan, and take the time to make sure that you're secure. At the very least, make sure that OT systems are NOT exposed to the Internet, and that they are isolated from your IT systems. Then make sure you keep things updated, including access and boundary control devices. Don't be a Shodan search away from a bad day.

9. Google hit with €403 million GDPR fine over location tracking

   Summary: Ireland’s Data Protection Commission (DPC) has fined Google €403 million (US $462m) for violations of the European Union's data privacy rules. Specifically, Google failed to process users' Web & App Activity and Location History data in compliance with standards established by the General Data Protection Regulation (GDPR). DPC has ordered Google to bring their data processing practices into compliance with GDPR within six months. The finding was the result of an inquiry that looked at Google's data processing practices between May 25, 2018 and February 4, 2020. DPC is Lead Supervisory Authority for Google in the EU because Google's European headquarters in in Dublin. A statement from Google says, in part, that "From 2019 onwards, [they]’ve significantly evolved [their] practices and launched robust tools that make managing location data simple."

   Lee's Take: This is a reminder that you can be held accountable for any past shortcomings in following GDPR. All you can do is make sure that you're fully following current guidance and documenting your decisions in that regard. If you've made improvements, document those. Make sure that you and your legal counsel are aligned; you've got attorney-client privileges, use them to your advantage.

Sam Bowne

1. AI CEOs Are Broke, Desperate and Lying About Doomsday

   The industry is telling two stories simultaneously. Story one told to the regulators and the public: this technology is so dangerous it could end civilization.

   Story two told to investment banks and credit agencies: this technology is so valuable that pension funds should buy our bonds even though they lose $20 billion a year.

   Both stories serve the same companies. The extinction narrative creates regulatory barriers that protect incumbents from open-source competition. The growth narrative extracts capital from institutional investors.

2. Stopping The AI Safety Cult ft. Adam Becker & Cal Newport

   Silicon valley leaders are in cults like Effective Altruism and Ray Kurzweil's Singularity Institute, believing extreme self-aggrandizing fictions about their own importance. They think AI and robots will create an amazing utopia with billions of future humans relying on them, so racing to develop AI as quickly as possible is their moral duty. Also, they are the only ones who can be trusted to do it ethically, so they need to win the race--they need to gather money and power as rapidly as possible to stop other, evil, people from getting there first.

   This is just the latest version of older ideas like the rich are chosen by God, with the right to rule. Elon's cult adds eugenics to it: his genes are superior to others so he needs to breed as much as possible. Peter Thiel says anyone who would slow the development of AI is the Antichrist.

   Our government leaders and other people outside the cult misuinderstand the insane apocalyptic statements they make about ending the world as real possibilities, but they are just as far from reality as Mike Huckabee's efforts to feed war in Israel to bring Jesus Christ back.

3. Attacker Hijacks AI Coding Assistant Session, Spreads Shai-Hulud Across About 100 Repositories

   Mandiant says an attacker hijacked an active AI coding-assistant session at an unnamed software-as-a-service provider and later spread Shai-Hulud across about 100 internal code repositories. Before the repository spread, the assistant recommended software that the attacker had poisoned, and the recommendation was accepted. The worm stole repository secrets and source code for the company's products.

4. Most Firms Unable to Recover Quickly from Ransomware

   Only four out of more than 800 clients (0.5%) assessed by Fenix24 came close to their own 24 to 48-hour ransomware recovery targets, and then only for partial business operations. None reached full operational capacity until several weeks after the incident. Recovery plans failed the same way each time, it said, looking sound on paper and coming apart once an attacker was inside.

5. Exclusive: US military had close call after using AI for false intelligence report, sources say

   An intelligence report said a Chinese ship in the Middle East was transporting components of a nuclear weapons program. But a chatbot the analyst had used inaccurately identified the material the ship was carrying. The report was “entirely false,” but it “almost started a war.”

6. An undercover Google analyst infiltrated a notorious supply-chain hacking gang

   Google had a mole inside TeamPCP’s inner circle since March 2025. TeamPCP has been running a huge supply-chain attack that started this Spring when they infected the open-source security scanner Trivy, the AI application programming interface tool LiteLLM, infrastructure of the web application security firm Checkmarx, the web app library TanStack, and the enterprise AI platform Mistral AI.

7. ‘Zuckoff’ App Detects Meta Smartglasses, as Meta Plans Camera-Free Version, Loses Money, and Offers Social Media Subscriptions (nypost.com)

   Zuckoff detects Bluetooth signatures broadcast by smart glasses and can estimate their distance.