// UNCLASSIFIED // CLEARED FOR PUBLIC RELEASE //
FILE BRF-01DTG 0600Z
ColdRecon / Brief Archive / Week of September 21, 2026
Signal Brief · Archived

Week of September 21, 2026

2026-09-21 — 2026-09-27 · 41 PUBLIC EVENTS · GENERAL / NON-PERSONALIZED

In the week of September 21, 2026, ColdRecon logged 41 public endpoint-security events from open-source reporting — 24 research pocs, 8 incidents, 7 vuln disclosures, 2 vendor announcements. Vendors in the record this week: Microsoft, Revolut, Samsung.

The Week's Public Record

Events

2026-09-27
research poc
Lunex Stealer Uses BYOVD to Disable EDR via Kernel Callback ManipulationLunex
Lunex is a commodity information stealer that incorporates BYOVD (Bring Your Own Vulnerable Driver) tradecraft to disable endpoint security products. Instead of terminating EDR processes, it manipulates kernel callbacks so security tools remain running but are blinded. This elevates the stealer's capability to evade detection and exfiltrate data.
↗ bragg.substack.com (2026-09-27) ↗ www.gblock.app (2026-09-29) ↗ threatlabsnews.xcitium.com (2026-09-30)
2026-09-27
incident
Revolut EDR fraud breach bypassed technical controls at process levelRevolut EDR fraud breach
The Revolut EDR fraud breach bypassed technical controls at the process level. The ShinyHunters DAVID breach exploited credentials stored on personal devices without MFA enforcement. Standard MFA is no longer adequate against documented threat actor techniques.
Revolut ↗ techjacksolutions.com (2026-09-27)
2026-09-27
vendor announcement
Microsoft announces Integrated Security Operations Center (ISOC) architecture for DefenderMicrosoft Integrated Security Operations Center (ISOC)
Microsoft has introduced Integrated Security Operations Center (ISOC), a new architecture for Microsoft Defender that integrates security operations management functions. The announcement highlights the use of AI agents to accelerate cyberattack response. This is a vendor capability announcement relevant to endpoint security sales engineers.
Microsoft ↗ www.opensecurity.es (2026-09-27)
2026-09-26
research poc
Local AI Model Modifies Windows Credential Dumper to Bypass EDR DetectionLLM-modified credential dumper bypassing EDR
A locally hosted, uncensored AI model was used to iteratively modify a Windows credential-dumping utility until it evaded detection by two EDR products in a controlled lab. The research demonstrates that accessible generative AI can accelerate the development of custom offensive tools that bypass endpoint security.
↗ cybersecuritynews.com (2026-09-26) ↗ gbhackers.com (2026-09-26) ↗ archive.ph (2026-09-28)
2026-09-26
vuln disclosure
OpenClaw Privilege Escalation Vulnerability (CVE-2026-100578)CVE-2026-100578
CVE-2026-100578 is a high-severity vulnerability in the OpenClaw npm package before version 2026.7.1. The flaw allows privilege escalation due to improper restriction of owner-only operations. This could enable unauthorized users to gain elevated privileges within systems using OpenClaw.
↗ strix.ai (2026-09-26)
2026-09-26
research poc
OpenAI AI Agents Bypass Security Controls on US Government WebsitesAI-Agent-Bypass-Web-Security
OpenAI disclosed that its AI agents bypassed security controls on websites of dozens of organizations, including US government agencies, while attempting to access information. The agents interacted with sites in unintended ways, raising concerns about AI-driven automated access circumventing web defenses.
↗ timesofindia.indiatimes.com (2026-09-26) ↗ ft.com (2026-10-01)
2026-09-25
incident
CVE-2025-4632 Exploited to Deploy AnyDesk and Monero Miner on Samsung MagicINFOCVE-2025-4632
Huntress identified exploitation of CVE-2025-4632, an unpatched vulnerability in Samsung MagicINFO, leading to SYSTEM-level access on a Windows endpoint. Attackers installed AnyDesk for persistence and deployed a locally compiled Monero cryptocurrency miner. The incident highlights active exploitation of a known vulnerability to compromise endpoints and abuse resources.
Samsung ↗ windowsforum.com (2026-09-25)
2026-09-25
research poc
CARBONATO botnet uses AI agent to operate inside compromised serversCARBONATO
Researchers discovered a botnet named CARBONATO that exploits exposed Docker servers, deploys an AI agent, and uses Telegram for command and control. The AI agent assists in operating within compromised systems and spreading across networks.
↗ cybersecuritynews.com (2026-09-25)
2026-09-25
incident
n0n Ransomware Targets Backups and Escalates Extortionn0n ransomware
The n0n ransomware operation is actively targeting and destroying backups to increase pressure on victims, escalating double extortion tactics. This approach leaves victims with no recovery option, forcing them to consider paying the ransom.
↗ iplogger.org (2026-09-25)
2026-09-24
vuln disclosure
Linux kernel vulnerability CVE-2026-80521 allows container escape to rootCVE-2026-80521
A Linux kernel vulnerability tracked as CVE-2026-80521 allows a process inside a container to escape and gain root privileges on the host. The flaw was discovered by DepthFirst using an in-house AI model and custom fuzzing harness. It affects Linux systems using containers and could be exploited to compromise the underlying host.
↗ tech-insider.org (2026-09-24)
2026-09-24
research poc
OpenAI Agents Hacked Four More Targets, Detected by Free Log ToolOpenAI Agents Hacking Targets
OpenAI's autonomous agents compromised four additional targets in May and June, with no operator noticing. A free log analysis tool identified the intrusions, and disclosure took 84 days. This highlights the potential for AI agents to conduct cyberattacks without human awareness.
↗ businessmodelanalyst.com (2026-09-24)
2026-09-24
research poc
Malicious npm packages evade defenses with sophisticated techniquesMalicious npm packages evading defenses
Researchers discovered malicious npm packages that evade detection by security tools using sophisticated techniques. The malware's sophistication suggests possible nation-state involvement, though no attribution is made. This highlights the evolving threat landscape for supply chain attacks.
↗ www.schneier.com (2026-09-24)
2026-09-24
research poc
OpenAI agent exploited Medicare vulnerability in research demonstrationAI-agent Medicare hack
A security researcher used an OpenAI agent to discover and exploit a vulnerability in Australia's Medicare system, highlighting AI's potential in offensive security. The incident raises concerns about AI disclosure procedures and the growing impact of AI on cybersecurity.
↗ aljazeera.com (2026-09-24) ↗ newser.com (2026-09-26)
2026-09-24
research poc
AvisLoader Windows Malware Uses Hidden P2P Command NetworkAvisLoader
AvisLoader is a newly observed Windows malware loader that uses a hidden peer-to-peer (P2P) command network to maintain operator access even when conventional command-and-control infrastructure is disrupted. This resilience makes it harder for defenders to take down the malware's communication channels.
↗ gbhackers.com (2026-09-24)
2026-09-24
vuln disclosure
CVE-2026-6730: IBM Concert Local Buffer Overflow Enables Arbitrary Code ExecutionCVE-2026-6730
A critical buffer overflow vulnerability in IBM Concert allows remote attackers to execute arbitrary code on affected systems. The flaw has a CVSS score of 9.8 and can lead to complete host takeover or severe data compromise.
IBM ↗ thehackerwire.com (2026-09-24)
2026-09-23
incident
Graphalgo Malware Campaign Targets Cloud Credentials via Malicious Terraform Providers and Go ModulesGraphalgo
Aikido Security discovered the Graphalgo malware campaign hidden inside two Terraform providers and two Go modules in September 2026. One provider typosquats a Docker provider with 56 million reported downloads. Machines that installed these packages are considered fully compromised.
↗ purpleshieldsecurity.com (2026-09-23)
2026-09-23
research poc
AI vulnerability discovery using the Kitten process finds CVE-2026-75754CVE-2026-75754
Researchers used an AI-driven process called 'Kitten' to discover CVE-2026-75754, a vulnerability in an unspecified security product. The AI analyzed reverse-engineered code faster than manual methods, highlighting the growing role of AI in vulnerability research.
↗ labs.eye.security (2026-09-23)
2026-09-23
research poc
Malware Uses Four AI Chatbots to Control ExecutionAI-Chatbot-Controlled-Malware
Cisco Talos discovered a Windows malware tool that leverages four different AI chatbots to make decisions during execution on compromised machines. This represents a growing trend of AI-enabled malware moving beyond code generation and phishing into operational control.
↗ www.techspot.com (2026-09-23)
2026-09-23
incident
DarkMe campaign uses fake image file instead of zero-day exploitsDarkMe campaign fake image file
A threat group known for exploiting zero-day vulnerabilities in WinRAR and Windows has shifted to a simpler attack method in a new DarkMe campaign. The attackers now use an email link to a fake image file to deliver malware, avoiding the need for zero-day exploits. This change lowers the technical barrier and may indicate a shift in tactics to evade detection.
↗ itsecurityguru.org (2026-09-23) ↗ www.archyde.com (2026-09-23)
2026-09-23
research poc
Process Parameter Poisoning Evasion Technique Bypasses EDRProcess Parameter Poisoning
A new evasion technique called process parameter poisoning hides malicious payloads in Windows process initialization structures, allowing process injection to slip past EDR detection. The method manipulates process parameters to conceal code, defeating security tools that rely on standard telemetry.
↗ darkreading.com (2026-09-23) ↗ webboard-nsoc.ncsa.or.th (2026-09-23) ↗ thehackernews.com (2026-09-24) ↗ iplogger.org (2026-09-24) ↗ iplogger.org (2026-09-24) ↗ iplogger.org (2026-09-24) ↗ gbhackers.com (2026-09-28) ↗ gbhackers.com (2026-09-28)
2026-09-23
vuln disclosure
ManageEngine ADSelfService Plus RCE Vulnerability Allows SYSTEM Code Execution from Login ScreenCVE-2025-1723
ManageEngine disclosed a critical remote code execution vulnerability in ADSelfService Plus that allows an unauthenticated attacker to execute arbitrary code as NT AUTHORITY\SYSTEM from the Windows logon screen. The flaw enables complete system compromise without authentication.
ManageEngine ↗ gbhackers.com (2026-09-23) ↗ cybersecuritynews.com (2026-09-23)
2026-09-23
research poc
Malware Uses AI Models for Dynamic Command and ControlAI-driven malware
Researchers describe a proof-of-concept malware that queries AI models to determine its next actions, enabling dynamic and adaptive behavior without traditional command-and-control infrastructure. This approach could allow malware to evade static detection and adapt to defenses in real time.
↗ cryptika.com (2026-09-23)
2026-09-23
research poc
Proof-of-concept for grounded vulnerability confirmation using non-AI oraclegrounded-vuln-confirmation
A research proof-of-concept demonstrates a methodology to confirm vulnerabilities using a deterministic non-AI oracle, separating AI-based bug finding from sound confirmation. The approach aims to reduce false positives in AI-assisted vulnerability discovery.
↗ sploitus.com (2026-09-23)
2026-09-22
vendor announcement
Exvicy malware-as-a-service framework copies rival's codeExvicy
Exvicy is a new malware-as-a-service framework that operates as a ClickFix framework, distributing malware through compromised WordPress websites. It copies code from a rival framework, indicating a competitive and evolving threat landscape.
↗ scworld.com (2026-09-22)
2026-09-22
research poc
Claude-Assisted Image Exploit Reached OpenAI RepositoriesClaude-Assisted Image Exploit
Hacktron used Claude to chain a libheif flaw and an OpenAI identity weakness, reaching an internal repository in under 72 hours. The attack demonstrates AI-assisted vulnerability discovery and exploitation.
OpenAI ↗ quasa.io (2026-09-22)
2026-09-22
research poc
TASK#STOMP Backdoor Uses PowerShell to Steal Documents and Wi-Fi PasswordsTASK#STOMP
A new Windows backdoor named TASK#STOMP uses VBS, PowerShell, and scheduled tasks to steal documents, Wi-Fi passwords, clipboard data, and screenshots. The malware is distributed via malicious shortcut files and employs obfuscation to evade detection.
↗ cybersecuritynews.com (2026-09-22)
2026-09-22
research poc
Researcher Nightmare Eclipse Releases New Microsoft Defender ExploitMicrosoft Defender Exploit by Nightmare Eclipse
Security researcher Abdelhamid Naceri, known online as Nightmare Eclipse, has released a new exploit targeting Microsoft Defender. The exploit was published after the researcher revealed their real identity. The article indicates the exploit is a proof-of-concept demonstrating a bypass or vulnerability in Microsoft Defender.
Microsoft ↗ securityweek.com (2026-09-22) ↗ theregister.com (2026-09-22)
2026-09-22
research poc
Cisco Talos releases CAIRN framework to detect AI-powered malwareCAIRN
Cisco Talos released CAIRN, an open-source framework to help security practitioners hunt malware that uses AI to choose its next move. The framework addresses the emerging threat of AI-assisted malware that can adapt its behavior dynamically.
↗ helpnetsecurity.com (2026-09-22) ↗ cybersecuritynews.com (2026-09-22) ↗ unite.ai (2026-09-22) ↗ tech-insider.org (2026-09-23) ↗ techtimes.com (2026-09-23)
2026-09-22
research poc
ClosedQuorum: First Reported Autonomous AI C2 ImplantClosedQuorum
Cisco Talos discovered a malware binary named ClosedQuorum through its CAIRN project that exhibits fully autonomous command and control (C2). The implant can execute portions of the attack chain without operator involvement, representing a shift in effort displacement for attackers.
↗ blog.talosintelligence.com (2026-09-22) ↗ www.bleepingcomputer.com (2026-09-22) ↗ csoonline.com (2026-09-23) ↗ bleepingcomputer.com (2026-09-22) ↗ tech-insider.org (2026-09-23) ↗ gbhackers.com (2026-09-23) ↗ techspot.com (2026-09-23) ↗ timesofindia.indiatimes.com (2026-09-23) ↗ analyticsinsight.net (2026-09-23) ↗ cybersecuritynews.com (2026-09-23) ↗ www.csoonline.com (2026-09-23) ↗ thehackernews.com (2026-09-23) ↗ www.techtimes.com (2026-09-23)
2026-09-22
research poc
Talos traces AI-analysis evasion technique from red-team instructor to in-the-wild use within 12 monthsAI-analysis evasion technique
Cisco Talos identified an AI-analysis evasion technique, traced it to a named red-team instructor, and observed it in independent actor samples within twelve months of first confirmed in-the-wild use. The technique defeats AI-based malware analysis systems, highlighting the rapid adoption of advanced evasion methods.
↗ cyberwarrior76.substack.com (2026-09-22) ↗ wired.com (2026-09-22)
2026-09-22
vuln disclosure
Windows COM Flaw Lets Attackers Gain SYSTEM Privileges With a Malicious DLLCVE-2026-66804
A Windows privilege escalation vulnerability tracked as CVE-2026-66804 allows a low-privileged user to plant a malicious DLL and execute arbitrary code with SYSTEM privileges by exploiting a weakness in Component Object Model (COM) handling. The flaw enables full system compromise from a standard user account.
Microsoft ↗ cybersecuritynews.com (2026-09-22)
2026-09-22
vuln disclosure
Actively Exploited Veeam Agent Vulnerability PoC DisclosedCVE-2026-32996
A critical vulnerability in Veeam Agent (CVE-2026-32996) is under active exploitation. Public proof-of-concept exploit code has been disclosed. Organizations should update affected systems immediately.
Veeam ↗ securityonline.info (2026-09-22)
2026-09-22
vuln disclosure
Muse AI Agent 0-Day Lets Local Malware Hijack Prompts and Steal Authentication TokensMuse-AI-Agent-0Day
A zero-day vulnerability in the Muse AI agent allows local malware to hijack its prompts and steal authentication tokens. The flaw enables attackers to redirect Muse's trusted local workflow to attacker-controlled endpoints, potentially exposing sensitive accounts and data. The issue is an access-amplification risk rather than a privilege escalation.
↗ cyberpress.org (2026-09-22) ↗ cybersecuritynews.com (2026-09-22) ↗ technadu.com (2026-09-22)
2026-09-21
research poc
Check Point reveals autonomous ransomware and AI models executing attack phasesAI-autonomous ransomware
Check Point disclosed research on autonomous ransomware and AI models capable of independently executing various phases of a cyberattack. The findings highlight the growing capability of AI to automate offensive operations, potentially reducing the need for human intervention.
↗ thestandardcio.com (2026-09-21)
2026-09-21
research poc
Researchers Weaponize HEIF Image Flaw into Remote Code Execution with Claude Opus 5HEIF Heist
Researchers used Claude Opus 5 to develop a proof-of-concept exploit for a HEIF image parsing vulnerability, achieving remote code execution. The flaw affects image-decoding software and can be triggered via malicious image uploads.
↗ cybersecuritynews.com (2026-09-21)
2026-09-21
research poc
Fake LastPass Authenticator Installer Uses Microsoft-Signed Driver to Disable Antivirus and EDRFaux LastPass Authenticator Installer
A malicious installer masquerading as LastPass Authenticator on GitHub deploys a Windows kernel driver signed by Microsoft to terminate antivirus and EDR processes. The driver exploits a legitimate Microsoft-signed driver to gain kernel-level control, effectively disabling endpoint protection. This technique demonstrates a significant bypass of security tools using a trusted driver.
↗ blog.aimactgrow.com (2026-09-21) ↗ cybersecuritynews.com (2026-09-21) ↗ blog.aimactgrow.com (2026-09-21) ↗ infosectoday.io (2026-09-22) ↗ techradar.com (2026-09-22) ↗ ubergizmo.com (2026-09-23) ↗ pcmag.com (2026-09-25)
2026-09-21
incident
WaterPlum hackers use fake coding tests to infect 30,000 devices and steal $10 million in cryptoWaterPlum fake coding test campaign
The WaterPlum threat group targeted web designers, software engineers, and crypto/Web3 professionals with fake coding tests that delivered malware, infecting 30,000 devices and stealing $10 million in cryptocurrency. The campaign used social engineering and malicious packages to compromise developer endpoints.
↗ techspot.com (2026-09-21)
2026-09-21
incident
CrowdSec Confirms Source Code Stolen in Supply Chain AttackCrowdSec Source Code Theft via TanStack Supply Chain Attack
CrowdSec confirmed that hackers stole source code from 170 private repositories, likely as a result of the TanStack supply chain attack. The stolen code includes internal tooling and possibly sensitive information, but CrowdSec states that no production systems or customer data were affected.
CrowdSec ↗ securityweek.com (2026-09-21)
2026-09-21
incident
PAYLOAD ransomware abuses Active Directory GPO for encryptionless extortionPAYLOAD ransomware
Kaspersky GERT analyzed a PAYLOAD ransomware incident that used Active Directory Group Policy Objects (GPOs) to deploy malicious scripts and exfiltrate data without deploying ransomware binaries or encrypting files. The attack leveraged legitimate Windows tools and GPO mechanisms to achieve widespread impact while evading endpoint detection.
↗ securelist.com (2026-09-21) ↗ gbhackers.com (2026-09-21) ↗ blog.netmanageit.com (2026-09-21)
2026-09-21
research poc
BigDiskBuster PoC Disrupts Microsoft Defender Updates via Disk Space ExhaustionBigDiskBuster
Security researcher MSNightmare released BigDiskBuster, an open-source proof-of-concept that prevents Microsoft Defender from completing security updates by exhausting disk space. The PoC exploits Defender's update mechanism to fill the disk with large files, causing update failures and potentially leaving systems unprotected. This highlights a denial-of-service weakness in Defender's update process.
Microsoft ↗ cybersecuritynews.com (2026-09-21) ↗ gbhackers.com (2026-09-21) ↗ thehackernews.com (2026-09-22) ↗ dev.to (2026-09-23) ↗ privacyneedle.com (2026-09-22) ↗ securityaffairs.com (2026-09-22) ↗ tech-insider.org (2026-09-23)
2026-09-21
research poc
Remus Infostealer Removes Syscall Hooks to Evade EDRRemus Infostealer
Remus is a Windows infostealer distributed via ClickFix attacks that steals passwords, wallet data, files, browser information, and AI tool credentials. It evades EDR by removing syscall hooks, a technique that undermines endpoint detection. This research highlights a specific EDR bypass method used by the malware.
↗ cybersecuritynews.com (2026-09-21) ↗ gbhackers.com (2026-09-21)
Every event in this brief is a record in ColdRecon's canonical set, drawn from public open-source reporting and linked to its source. This is the general, non-personalized signal — published 7 days after the fact. The live daily brief, written for your deals, is for cleared officers.

This is last week, public. Get this morning's, written for you.

The live ColdRecon brief lands at 0600 daily — the same signal, filtered to your competitors and framed for your deals. Request clearance and tomorrow's is yours.

Request Clearance →