Agents are no longer a subject of security research. They’re doing the attacking, they’re doing the defending, and they’re the reason a messaging app is blocking its own screenshots.

The agents are doing the attacking now

The story of the day is Amazon Threat Intelligence’s disclosure from yesterday: autonomous AI agents running on Cairn, an open-source penetration-testing framework, found and exploited a zero-day in MONAI Label, a medical imaging server, and took root on a university server. The flaw is CVE-2026-133968 (https://thehackernews.com/2026/11/attackers-ai-agents-found-zero-day-in.html), a path-traversal bug in the image upload function, scored 9.8. The campaign ran about 48 hours in September against East Asian university servers. The operator gave the agents a target list and one instruction — obtain shell access — and they picked the methods themselves. From mapping the API to arbitrary file write took about eleven minutes; root came in a later session, through a reverse shell planted in a file Python loads at startup.

Two things stay with me: how little direction the agents needed — a goal and a target list — to find a zero-day nobody knew about and weaponize it. Second, the state of the fix: merged October 19, still no patched release — the newest version is 0.8.5, from 2024 — and authentication is off by default, so anyone who can reach a server can exploit it. Amazon found the whole thing only because the operator left the Cairn control panel exposed with no password, and attributes the campaign to a China-linked actor with moderate confidence. The patch exists. The release doesn’t. That gap is the whole story.

Back in June, Google DeepMind put up to $10 million behind multi-agent AI safety research (https://www.technologyreview.com/2026/06/11/1138794/google-deepmind-is-worried-about-what-happens-when-millions-of-agents-start-to-interact/), with Schmidt Sciences, the UK’s ARIA, the Cooperative AI Foundation, and Google.org — a grant call for research into what happens when millions of AI agents start taking instructions from each other: prompt injections that spread on their own, self-guiding malware, attacks hopping from agent to agent at machine speed. Rohin Shah put it plainly: “there isn’t really a field of research for multi-agent safety yet.” Five months later, the agents that field was meant to study found a zero-day in eleven minutes. The funding was early. The threat wasn’t late.

The defense is adapting in ways that would have sounded absurd a year ago. Signal announced yesterday that it is blocking screenshots by default (https://signal.org/blog/protecting-your-privacy/) — on iOS 27, screenshots are blocked inside the app, with a new “Block Siri AI” setting; on Android, its Screen Security turns on when it detects an AI assistant trying to read the screen. Same move Signal made last year against Microsoft Recall. Whittaker is blunt that the remedy is incomplete: it doesn’t stop agents reaching content already on the device, or third-party agents that arrive with broad permissions. The encryption still holds. The screen is the new leak — and the agent reading it ships inside the phone.

The patch queue

CISA added four flaws to the exploited list yesterday; three were covered here before. The fourth is new to this feed, and the most instructive of the batch. CVE-2025-42884 is an unauthenticated JNDI injection in SAP NetWeaver’s Enterprise Portal (https://cyberpresso.com/blog/sap-netweaver-portal-cve-2025-42884-cisa-kev-jndi): an attacker with no login can inject JNDI environment properties or a URL used during a lookup, and disclose or modify server information. CISA gave federal agencies until November 23. The timeline is the point: SAP patched this a year ago, in November 2025, and scored it 6.5, medium. Onapsis, which reported it in August 2025, says the same flaw could reach an attacker-controlled registry and, with the right gadget chain, remote code execution. A medium from last year is this week’s exploited list. Scores expire. Exposure doesn’t.

Apache disclosed a critical Tomcat flaw this week, CVE-2026-133325 (https://cybersecuritynews.com/critical-apache-tomcat-flaw/), scored 9.8: in clusters secured with TLS 1.3 and a pre-shared key, Tomcat doesn’t validate the peer, so an unauthenticated node can join a trusted cluster. The honest part of the advisory is the fix status: the patched versions — 11.0.30, 10.1.63, 9.0.125 — are not released yet. Until they are, Apache’s advice is to switch cluster communication to TLS 1.2 with the pre-shared key kept on. No exploitation has been reported, and Apache doesn’t claim joining a cluster gives remote code execution. Encrypted but unauthenticated is a locked door nobody checks the key for. Don’t wait for the release to make the config change.

Synology patched a critical flaw in DiskStation Manager and BeeStation OS, CVE-2026-120382 (https://socprime.com/blog/cve-2026-120382-critical-synology-dsm-oauth-flaw-enables-arbitrary-file-access/), scored 9.8: an argument-injection bug in the OAuth API that lets an unauthenticated remote attacker read or write arbitrary files and cause denial of service. Fixed builds are out for DSM 7.4 and 7.3; fixes for the 7.2 branches, 7.1, and BeeStation are still ongoing — interim advice is to deactivate the OAuth Service package. Synology hasn’t published the endpoint, says no exploitation has been observed, and doesn’t claim remote code execution. The confirmed impact is file access and denial of service — which, on a NAS, is enough.

GnuPG disclosed a critical memory bug in GPGME, CVE-2026-129975 (https://thehackernews.com/2026/11/gnupg-warns-critical-gpgme-flaw-could.html): a double free triggered by an OpenPGP key with a crafted very long user ID. Two routines that read status lines and key listings back from gpg move data to a bigger buffer but only record the new address if they finish cleanly; an early return leaves a pointer to memory already freed. The project says it can exfiltrate data from applications by sending signed data, and remote code execution can’t be ruled out. The fix is GPGME 2.3.1, with Gpg4win 5.1.2 for Windows; disclosure was pulled forward from November 25 because the Ruby2Shell bug forced GitLab to update early. The bug is in the library underneath the app — mail clients, Flatpak, Samba, Gpg4win all link against it — so check the library version, not just the app version.

WhatsApp’s November advisory fixes three flaws; the one to read first is CVE-2026-23871 (https://gbhackers.com/whatsapp-security-flaw/). A crafted Click to WhatsApp ad-reply message could make an Android phone fetch media from an attacker-controlled URL and expose the user’s IP — no tap needed, just rendering the message. WhatsApp says it “may have been abused in conjunction with specific spam messages”; its own security team found it while disrupting a scam campaign. The second flaw is worse on paper: a SQL injection, CVE-2026-102769, where tapping a crafted “order details” message from an unofficial client could delete locally stored files. Android 2.26.35.74 covers all three. The scam angle is the tell — the flaw made fraudulent messages look more credible, which is exactly what a scam campaign needs.

The breach ledger

The FBI is investigating an apparent breach of its National Crime Information Center database, CNN reported yesterday (https://www.reuters.com/world/us/fbis-criminal-information-database-hacked-cnn-reports-2026-11-21/), citing two people briefed on it. NCIC is the system federal, state, and local law enforcement use to share warrants and background checks, millions of times a day. It’s the second intrusion involving an FBI system in recent months; in September, ShinyHunters hit an FBI jobs website and took employee personal data. There’s no public confirmation of what was taken or altered, and no attacker has been identified. Be careful with the word: apparent, suspected, under investigation — not confirmed. But the shape of it is what matters. The system that tells an officer whether the person in front of them is wanted is the one system you least want in doubt, and the doubt is now public.

The Carnegie Endowment for International Peace disclosed that a breach at Soapbox Engage, its online-donations vendor, exposed contact and donor information (https://www.nextgov.com/cybersecurity/2026/11/carnegie-endowment-discloses-breach-exposing-contact-and-donor-information/417378/). An unauthorized party got into the vendor October 28 to 30 and reached Carnegie’s donor management database — names, email addresses, phone numbers, and for donors, giving histories. Full credit card information wasn’t exposed. Carnegie calls the affected group a “small percentage,” gives no count, and notified everyone in the database as a precaution. The data itself is modest. The use of it isn’t: giving histories let a phisher write to you as Carnegie, referencing what you gave last year, and sound exactly right. A donor list is a phishing list with the targeting done for you.

Triple Whale, the e-commerce analytics provider, disclosed yesterday that a threat actor had unauthorized access to some of its internal systems (https://www.geeknative.com/244332/dungeons-dragons-partner-sneak-energy-reports-customer-data-breach-starting-on-launch-day/). Detected October 31, contained November 1, disclosed November 20. Sneak Energy’s customer notice fills in the shape: store data accessed October 30 to November 1 — customer names, email addresses, phone numbers, mailing addresses, and total amounts spent. No payment details, no order details, no passwords. Sneak says it’s one of more than 65,000 companies affected. The breach is at the analytics vendor, but the data is your customers’ — and the system holding the customer list is never the one you think of.

Rokkatei, the Hokkaido confectionery maker, disclosed that its online shop’s payment page was tampered with between October 21 and November 18. About 8,000 card entries are affected (https://smartscope.blog/en/blog/rokkatei-online-shop-potential-credit-card-information-leak-2026/) — full card numbers, names, expiration dates, and security codes — plus names, addresses, phone numbers, and emails. One fraudulent charge of about ¥50,000 has already been confirmed. The company doesn’t store card data at all; the tampered page intercepted the details as customers typed them. “We don’t store card data” is not the defense people think it is. The payment page is the storage.

The USDA spent this week taking its own systems offline after its Cybersecurity and Privacy Operations Center uncovered a security incident (https://www.nextgov.com/cybersecurity/2026/11/usda-grapples-security-breach-it-staff-take-multiple-systems-offline/417350/). An internal email Monday ordered a reset of all database passwords in all environments — which meant resetting every application: Farm Service Agency loan software, the ProTracts program that releases farmer payments, the Service Desk mapping tool. Farmers couldn’t check loan and aid statuses, and ARC/PLC sign-ups ahead of the December 11 deadline were disrupted. Who did it, and whether any data was taken, is unclear; the undersecretary says no personal data was lost and the protections worked. Both can be true at once: the protections held, and the farmers still couldn’t get paid. Resetting every password in every environment is the right call. It’s also the reason the outage lasted days.

The West Virginia Legislature identified a cybersecurity incident Thursday afternoon; its website was down Friday (https://westvirginiapulse.com/articles/west-virginia-legislature-cybersecurity-incident-website-down-nov-2026) while outside specialists helped investigate, and internal systems and remote access were shut down. An internal memo described it as a possible ransomware attack — but the spokesperson will only say “cybersecurity incident,” with no confirmation of ransomware, no actor named, and no word on whether data was taken. Don’t call it ransomware until someone who knows says so.

Equinor disclosed a cyberattack on its IT systems on the evening of November 17; the next day, Skagerak Energi disclosed one of its own (https://www.darkreading.com/cyberattacks-data-breaches/second-norwegian-energy-company-breached). Both said operations and production were unaffected, and neither reported signs of data theft. Equinor said Friday morning that its attack had been stopped, adding that it “does not appear to be specifically targeting Equinor.” Skagerak hasn’t said its attack is stopped. There’s no evidence the two are connected — but Norway’s National Security Authority had warned weeks earlier of high cyber risk, and the Police Security Service warned November 12 about cyber sabotage and Russian intelligence mapping Norwegian infrastructure. Two companies, two days, no link established. The pattern doesn’t need a link to be a warning.

Bell Canada confirmed that an intruder it calls an “agentic threat actor” broke in on November 17 (https://www.bleepingcomputer.com/news/security/bell-canada-says-agentic-hacker-stole-enterprise-customer-data/) and stole enterprise customer data, entering through a pre-production environment of Maximo, the system Bell uses to order and provision telecom circuits. Taken: business contact details, ticket titles, and some IP addresses tied to network elements. Not taken: credentials, highly sensitive personal information, PDFs, or attachments. Bell says it contained the incident immediately, services weren’t affected, and it’s sharing indicators of compromise privately with affected customers — not publicly. “Agentic” is Bell’s word, and it’s doing a lot of work: it suggests AI-powered tools were involved, but that’s an inference, not a fact. What is confirmed is that the IOCs aren’t public, so everyone else is hunting blind.

The tooling

Rapid7 published research this week on a custom SoftEther VPN bridge attributed to Webworm (https://www.rapid7.com/blog/post/tr-spying-on-governments-no-malware-custom-webworms-toolkit-softether-bridge/), a China-aligned group that targets government networks, with moderate confidence — a rebuilt SoftEther VPN Bridge, rebranded “NtMonitor VPN” and disguised as network monitoring software. Here’s the part worth sitting with: no backdoor, no remote code execution, no command shell. Its only job is to bridge a remote attacker-controlled network directly into the victim’s internal network — one compromised host becomes a gateway to everything behind it. Only one antivirus vendor flagged the public sample. Zero lines of malware — there’s nothing in it that looks like an attack, because the attack is a legitimate VPN doing exactly what a VPN does. If it looks like your own tooling, your tooling list is the detection.

Kaspersky’s research team documented AuditTeam (https://cybersecuritynews.com/auditteam-ransomware-gang/), a ransomware group that compromised dozens of organizations in at least seven countries between January and October — China, Finland, Hong Kong, Italy, the Philippines, Russia, and Senegal. Its tool is OpsAgent, a modular framework likely written with AI assistance, rewritten from C++ to Rust during the campaign. One add-on module loads a digitally signed antivirus driver and uses it to kill about 220 processes: security products, monitoring agents, backup services, databases. The ransom note is laid out like an audit log, addressed to executives and legal and compliance teams. The group says its mission is to educate organizations about cybersecurity. They call it an audit. The invoice is a ransom note — and the “audit” starts by killing the tools that would have caught it.

ThreatFabric disclosed Bazuka (https://thehackernews.com/2026/11/bazuka-android-malware-uses-fake-screen.html), a modular Android banking trojan found in late October during an NFC-relay investigation: a dropper, a core, and 23 modules — overlays, keylogging, 2FA and SMS interception, remote control, crypto-address swapping. The overlay trick is new: it screenshots your home or lock screen and displays that back to you, hiding what it’s doing underneath. The AI part is in the operator panel, where a language model scores victims so operators can prioritize the valuable ones at scale. Delivery came through a fake TikTok18+ adult-content app. The AI doesn’t steal the credentials. It picks who to steal from — and it’s the part that scales.

What I’d take from today

Three things, all the same thing from different angles. The agents have arrived as attackers: they found a zero-day in eleven minutes, and a telecom is blaming an “agentic” intruder. The agents have arrived as defenders too — Signal is blocking its own screenshots because the operating system itself is now the surveillance surface. And the oldest problems haven’t gone anywhere: a year-old medium on the exploited list, a NAS with an unauthenticated file-write flaw, a payment page skimmed for four weeks, a donor list that is a phishing list. The new threat doesn’t replace the old ones. It stacks on top of them.

Leave a Reply

Trending