Triaging the Week 144

Hello there πŸ‘‹

Welcome back to the Kraven Security weekly newsletter, triaging the week. We round up the week’s top news stories, highlight our featured article, give you some learning resources, and finish with a few personal notes about what’s happening at the company. Enjoy!


Top News Stories

Triaging the Week News Stories

Theme 1: AI Goes Operational

Stories

πŸ—žοΈ Threat Actors Deploy ARTEX for Autonomous Cybercrime (Amazon Threat Intelligence) – Criminals are using ARTEX, an open-source agentic AI framework, to automate credential theft and deepfakes. Stolen API keys go straight back into the framework, paying for more credential hunting and for real-time deepfakes that get past KYC liveness checks. πŸ”Ž Threat Hunting Package

πŸ—žοΈ Agentic AI Weaponization: A New Era of Machine-Speed Cyber Attacks (Google Threat Intelligence Group) – GTIG found an exposed “exploit_agent” framework that reads patches, writes exploits and runs them with no human involved. It’s the first fully autonomous pipeline from vulnerability discovery to working exploit seen in the wild.

πŸ—žοΈ Anthropic Unveils Claude Fable 5.2 & Mythos 5.2 (Anthropic) – Fable 5.2 is public and ships with safeguards. Mythos 5.2 scored 89.6% on the CVE Rediscovery benchmark and has no consumer misuse safeguards, so access is limited to the Cyber Verification Program.

πŸ—žοΈ Apple Overhauls macOS Security Controls and Permissions to Mitigate AI Agent Privacy Risks (Apple) – macOS 27.3 beta 2 splits up Full Disk Access, Device Control and Data Access permissions and adds menu bar indicators. Apple says the changes are aimed at the risks of autonomous AI agents.

Recommendations

β˜‘οΈ Audit and lock down every AI inference endpoint, model gateway and management interface. Rotate default credentials, remove hardcoded keys from repos, and bring AI apps under your zero-trust architecture.
β˜‘οΈ Watch cloud billing and AI service usage logs for unexpected spikes in API spend, which can mean hijacked keys.
β˜‘οΈ Update detection rules to catch the traffic and code traces left by multi-agent frameworks such as CAIRN, ARTEX and PentAGI.
β˜‘οΈ Use defensive AI where it helps, such as Fable 5.2 for code review in DevSecOps, and require a human in the loop for critical deployments in your AI governance policy.
β˜‘οΈ Apply least privilege on endpoints. Limit macOS Full Disk Access and automation permissions to trusted workflows.

Theme 2: The Shrinking Patch Window

Stories

πŸ—žοΈ Google Project Zero Slashes Disclosure Timelines: The End of the 90-Day Standard (Google Project Zero) – Project Zero is moving to a “Full X” disclosure policy that shrinks 90 days to 30 by 2028, with an interim “Full 60” deadline from 2 February 2027. The reason: LLMs make “zero patch-to-exploitation time” a realistic threat model.

πŸ—žοΈ Java4Shell (CVE-2026-114938): Critical OpenJDK RCE Vulnerability (Calif) – A malformed Base64 string ending in a stray = makes OpenJDK’s hand-written AVX-512 decoder calculate a negative length, which overwrites 64 bytes of heap. That gives unauthenticated RCE in Java 17, 21, 25 and 27 underneath Jira, Confluence, Jenkins, Keycloak and Kafka.

πŸ—žοΈ Critical NGINX HTTP/2 RCE Flaws Expose Web Servers to Complete Takeovers (Wiz) – Two CVSS 9.2 HTTP/2 RCE flaws (CVE-2026-123007 and CVE-2026-123633) have a public PoC and affect versions back to 1.9.5. One of them went unnoticed in the codebase for more than a decade.

πŸ—žοΈ PHP Patches 7 Critical Flaws Exposing Servers to Remote Code Execution (PHP Foundation) – The release fixes 17 flaws, 7 of them critical, and PoCs are already circulating. Many need non-default setups such as PHP-FPM or file-based sessions, but PHP runs almost everywhere, and versions 8.1 and older are end-of-life and won’t get fixes.

πŸ—žοΈ Microsoft Edge Introduces ‘Lockdown Mode’ to Defeat Zero-Day Exploit Chains (Microsoft Browser Vulnerability Research) – Lockdown Mode turns off WebGL and WebGPU. Microsoft estimates that would have blocked about 60% of in-the-wild privilege escalation exploits since early 2022.

Recommendations

β˜‘οΈ Patch now:

  • OpenJDK: the November 2026 updates.
  • NGINX: 1.30.7 or 1.31.9. Do a full restart, because a config reload won’t replace the running binary.
  • PHP: 8.6.0, 8.5.13, 8.4.28, 8.3.37 or 8.2.36. Move off PHP 8.1 and older.

β˜‘οΈ Where you can’t patch yet, use temporary mitigations:

  • Java: -XX:-UseBASE64Intrinsics
  • NGINX: http2 off;
  • PHP: turn off session upload progress and separate session storage.

Longer term, build non-patch controls such as feature toggles.

β˜‘οΈ Assume exploitation starts the day a patch ships. Shorten remediation SLAs for internet-facing systems before the Full 60 deadline on 2 February 2027.
β˜‘οΈ Keep a continuous inventory and run SCA to find Java runtimes and NGINX instances hidden in containers and appliances. Watch for NGINX worker crashes and unusually large HTTP/2 streams.
β˜‘οΈ Pilot Edge Lockdown Mode for high-risk users: Edge Canary, the #edge-sdsm-plus flag, and Strict mode. First, map out which internal tools depend on WebGL or WebGPU.

Theme 3: Developers in the Crosshairs

Stories

πŸ—žοΈ North Korean Hackers Exploit Deepfakes in Massive Crypto Heist (U.S. Department of Justice) – Deepfake video interviews and malware disguised as “technical fixes” led to more than $400M stolen and over 30,000 devices compromised. The campaign used AI-generated personas to manipulate people rather than exploit software flaws.

πŸ—žοΈ The Trigram Campaign: Fake Jobs and Real Infostealers Target Devs (Stormshield) – This campaign, likely North Korean, impersonates more than 50 companies and tracks victims on a custom back-office platform. Candidates are tricked into pasting a malicious command into their terminal to pass a fake “spam verification” check. πŸ”Ž Threat Hunting Package

πŸ—žοΈ Microsoft’s ‘agent-framework’ PyPI Package Compromised (Wiz) – Versions 1.26.1 and 1.26.2 drop a .pth file into site-packages. The credential stealer then runs every time Python starts, even if the package is never imported. πŸ”Ž Threat Hunting Package

πŸ—žοΈ Optimizely Supply Chain Attack Delivers Malware to 178 Major Brands (Report URI) – Malicious JavaScript ran on 178 brand sites, including Zoom, BBC and CNN, on 15–16 November. It swapped crypto wallet addresses on the page in real time, skimmed card data, and pushed the Coruna and DarkSword exploit kits to devices on iOS 18.6 or earlier. πŸ”Ž Threat Hunting Package

πŸ—žοΈ API Keys Stolen from Docker Hub Images in Under 8 Seconds (Stony Brook University) – Bots the researchers call “ContainerBots” pull secrets out of static image layers through registry APIs in as little as 7.86 seconds, without running the container. Runtime controls can’t stop this, and AWS, OpenAI and Anthropic keys were all compromised within 15 minutes.

Recommendations

β˜‘οΈ Brief hiring managers and engineers on these lures. Nobody should download software or paste terminal commands during an interview or a “verification” check, and recruiters and domains should be verified through a separate channel.
β˜‘οΈ Hunt for agent-framework 1.26.1 and 1.26.2 in lockfiles, CI logs, container images and virtualenvs, and pin to 1.26.0. If you find exposure, rotate every secret (Azure CLI, DevOps PATs, npm/PyPI tokens, Hugging Face credentials, .env files) and audit Azure identities and GitHub Actions OIDC.
β˜‘οΈ Treat any secret published in an image as compromised. Revoke and rotate it rather than just deleting the image, add secret scanning that fails CI builds, and switch to short-lived credentials injected at runtime.
β˜‘οΈ Protect against third-party script compromise with CSP and SRI. Optimizely customers should review their projects for injected scripts and token abuse. Affected users should update to iOS 18.7.3 or 26.3.
β˜‘οΈ Require phishing-resistant MFA and least privilege on code hosting and messaging platforms. Alert when a terminal process makes unexpected network requests.

Theme 4: Hiding in Plain Sight

Stories

πŸ—žοΈ Russian Threat Group COLD RELIC Turns Cloud Compromises into Phishing Springboards (Google Threat Intelligence Group) – This FSB-linked group uses residential proxies located near its targets to avoid “impossible travel” alerts. It registers rogue MFA devices, passkeys and backup codes so its access survives password resets.

πŸ—žοΈ ShadowCloud APT Leverages GitHub & SharePoint for Covert Command & Control (Kaspersky) – The China-linked GitDoor and GraphDoor backdoors send their C2 traffic through trusted GitHub and SharePoint APIs. Perimeter and signature-based controls can barely see it. πŸ”Ž Threat Hunting Package

πŸ—žοΈ ARMORY CASTLE Hackers Exploit VMware vCenter Using Custom ‘BONESAW’ Malware (Google Threat Intelligence Group) – This PRC-nexus actor chained two zero-days, including CVE-2026-123890 in vCenter. Its BONESAW tool abuses DCE/RPC to bypass SRP authentication, exfiltrate databases and forge admin tokens. πŸ”Ž Threat Hunting Package

πŸ—žοΈ 0% GPU Utilization: How SANSAR Hides Linux Cryptomining (Elastic Security Labs) – REF3364 gets in through exposed Ray job agents and hides its C2 behind Tor and Ethereum smart contracts. A modified NVML wrapper reports 0% GPU usage to admins while the GPUs mine. πŸ”Ž Threat Hunting Package

Recommendations

β˜‘οΈ Add residential proxy and VPN reputation data to authentication logs and enforce it with conditional access. When remediating an account, revoke all sessions, passkeys and MFA devices instead of only resetting the password. Add DLP to stop credentials being stored in cloud files.
β˜‘οΈ Set a baseline for GitHub and SharePoint/Graph API use and alert on unusual processes, user agents or repo access. Tighten egress so only approved workloads can reach even legitimate cloud endpoints.
β˜‘οΈ Patch vCenter for CVE-2026-123890 and restrict management ports 2012, 9080 and 5480. Extend EDR to Linux appliances to catch webshells such as SLIMSAVAGE and CAGEMATCH.
β˜‘οΈ Lock down AI/ML infrastructure:

  • Never expose unauthenticated Ray job APIs to the internet.
  • Run YARA rules for SANSAR and netns.
  • Alert on unauthorized DKMS builds.
  • Block Tor and unexpected blockchain queries from GPU clusters.

Theme 5: Government Strikes Back

Stories

πŸ—žοΈ 354 Arrested in Dark Web Cybercrime Crackdown (Europol) – Arrests covered Archetyp, Nemesis, Tor2Door, Bohemia and Kingdom. Seizures of more than $166M, 152 firearms and 4.8 tonnes of drugs show how closely dark web markets are tied to physical organised crime.

πŸ—žοΈ DOJ Dismantles Gootloader Infrastructure Tied to $29M in Ransomware Attacks (U.S. Department of Justice) – Operation Endgame dismantled Gootloader, which has infected more than 25,000 devices since 2018. Its access-as-a-service model enabled at least 44 US ransomware attacks and $29M in losses.

πŸ—žοΈ Scattered Spider Leader Sentenced to 6 Years for $11.6M Crypto Theft Scheme (BleepingComputer) – A 24-year-old British leader was sentenced for a smishing and SIM-swapping scheme. He is the fifth and final defendant in the case, another sign that law enforcement is catching up with English-speaking crime groups.

πŸ—žοΈ New DOJ & DHS Task Force Launched to Disrupt Global Cyber Scams Targeting Americans (The Record) – The new Threat Disruption Center has already taken down more than 1,600 sites and flagged over 6,600 wallets. Tech partners freeze payments and ban accounts under their terms of service while law enforcement pursues cases at the same time.

πŸ—žοΈ Record Cyber Incidents Reported to UK Regulator (The Record) – The ICO received a record 3,463 cyber incident reports, up 35% on the previous year.

πŸ—žοΈ NSA Classified Network Disruption Highlights Critical Infrastructure Fragility (The Record) – The NSA Net outage was officially put down to maintenance. It happened alongside aviation, banking and Microsoft failures, which led lawmakers to call for an independent investigation and shows how fragile critical digital supply chains are.

Recommendations

β˜‘οΈ Watch dark web and threat intel feeds for leaked credentials and data, since displaced actors often dump stolen assets. Check your logs against the IoCs law enforcement releases, including Gootloader’s.
β˜‘οΈ Replace SMS 2FA with phishing-resistant MFA such as FIDO2, and alert on unauthorised MFA device enrolments.
β˜‘οΈ Shut off common ways in. Filter web traffic to block compromised download sites, update awareness training for smishing and crypto investment scams, and use zero trust so a broker’s single foothold doesn’t turn into ransomware.
β˜‘οΈ Refresh incident response plans. Test out-of-band communications, and have a red team challenge the “it’s just maintenance” assumption when several outages happen at once.
β˜‘οΈ Financial, tech and crypto organisations should set up real-time intelligence sharing with the Threat Disruption Center. Update threat models for the instability that follows big takedowns.


Feature Video

Did Anthropic’s Glasswing or OpenAI’s Daybreak projects break vulnerability management?

Two numbers to keep in mind with all the hype πŸ‘€

πŸ‘‰ CVE records in 2026: 86,500+. That’s nearly 80% more than all of last year, with two months still left.

πŸ‘‰ Additions to CISA KEV: up only 11% (that is, vulnerabilities being exploited in the wild). Microsoft’s data shows that only about 0.5% of all CVEs are ever linked to real exploitation.

So most of this flood is never used in the wild.

But the part that is concentrated on unauthenticated, remote, edge devices, and weaponized before your triage cycle finishes!

πŸ“ˆ Verizon DBIR: edge devices and VPNs went from 3% to 22% of breaches that started with an exploited vulnerability.

The median workstation vulnerability gets patched in about 2 weeks. The median network appliance vulnerability takes more than a year.

The only way to combat this vulnerability explosion is to stop triaging by CVSS and ask three questions:

1️⃣ Is it on the edge?

2️⃣ Is it unauthenticated?

3️⃣ Is it in KEV, or being sold by access brokers?

You can’t patch 86,500 things. You can decide in advance which questions matter, and which ones to prioritize immediately.

Feature Course


Learning Resources

Triaging the Week Learning Resources

Cyber Training

Tools