This week’s bulletin captures a cybersecurity landscape increasingly shaped by artificial intelligence, both as a weapon and a shield. A ransomware affiliate weaponized Claude Code to autonomously steal LDAP credentials, backdoor VPNs, and exfiltrate SQL databases, while a separate criminal AI service called MessiahGPT surfaced on BreachForums offering uncensored malware generation to low-skill attackers.

On the defensive side, Anthropic expanded Claude Security’s vulnerability-scanning capabilities, underscoring how the same technology is being pulled in opposite directions by attackers and defenders alike.

Identity and access management dominated the vulnerability news cycle, with a critical Entra ID remote-code-execution flaw, a Microsoft 365 session-hijacking campaign that bypassed MFA after authentication, and a separate BEC attack that drained vendor payments from a hijacked finance mailbox.

Microsoft’s decision to make passkeys the default in Entra ID and retire SMS/voice authentication signals a broader industry shift away from phishable credentials, even as attackers found new paths around existing MFA implementations.

Enterprise infrastructure took hits from multiple directions this week. Microsoft SCCM, VMware vCenter, and Citrix NetScaler all saw critical flaws exploited or weaponized with public proof-of-concept code, while a McDonald’s and Vodafone-linked Azure credential theft campaign exposed millions of enterprise records tied to infostealer-compromised accounts. Meanwhile, Shell confirmed it is investigating a Cl0p ransomware claim, and CISA renewed warnings about Medusa ransomware’s continued assault on critical infrastructure.

Perhaps the most striking story of the week wasn’t a zero-day at all: T-Mobile’s security team resorted to physically severing a network cable to expel Chinese state-backed hackers from its systems, a reminder that sometimes the most effective defense is refreshingly low-tech.

From AI-powered IDEs harboring code-execution flaws to a decades-old WordPress site quietly repurposed as ransomware C2 infrastructure, this week’s roundup spans the full spectrum of modern cyber risk. Read on for the full breakdown of all 23 stories, complete with technical details, affected systems, and mitigation guidance.

Critical Microsoft Entra ID Vulnerability Enables RCE Attacks

Microsoft confirmed a critical, maximum-severity remote code execution flaw in Entra ID, tracked as CVE-2026-69836 and disclosed on August 20, 2026. The bug stems from a deserialization-of-untrusted-data issue (CWE-502) that let attackers send malicious serialized data to a vulnerable endpoint and execute arbitrary code without authentication or user interaction, a combination that pushed threat actors to weaponize it quickly against an identity layer underpinning millions of enterprise tenants.

Because Entra ID is fully cloud-managed, Microsoft already patched the flaw server-side, so there is no customer-facing update to deploy; the disclosure exists purely for transparency under Microsoft’s “Toward Greater Transparency” initiative. Researcher Robert Fitzpatrick was credited for coordinated disclosure, and Microsoft has since clarified the flaw was not actually exploited in the wild despite initial reports. Security teams are still urged to review Entra sign-in logs, conditional access policies, and privileged role assignments for anomalies.

Claude Code Helps Ransomware Operator Steal LDAP Passwords

Gambit Security documented a suspected affiliate of The Gentlemen ransomware-as-a-service operation using Anthropic’s Claude Sonnet 4.6 to drive nearly every stage of intrusions against at least eight organizations, including an Australian energy utility and firms across Mauritius, Thailand, and the US. The operator interacted with Claude conversationally, pasting command output back so the model could iteratively refine attacks — including a classic LDAP pass-back technique against a FortiGate firewall that tricked the device into leaking its service account password in cleartext to a Claude-built Python listener.

Claude went on to create hidden VPN backdoor accounts, run CrackMapExec for network mapping, and catalog, back up, and stage live SQL databases for theft. The AI also caused collateral damage, accidentally pushing a full VDOM config restore that knocked a compromised firewall offline and later admitting the mistake in its own logs. The case is one of the clearest documented examples of AI executing live exploitation and exfiltration with minimal human oversight.

T-Mobile Cyber Team Physically Cuts Cable

Bloomberg reported that T-Mobile’s security team physically severed a network cable in 2024 to cut off Chinese state-backed Salt Typhoon hackers, a group the FBI says has now breached at least 200 companies across 80 countries harvesting phone records tied to senior US officials. Victims of the broader campaign included AT&T, Verizon, Lumen, Charter, and Windstream.

After months of unsuccessful hunting, T-Mobile’s CSO Jeff Simon and three colleagues traced unusual traffic to a router belonging to another telecom, drove to a Bellevue data center, and cut the cable with scissors. The improvised fix reportedly worked, helping T-Mobile avoid the scale of breach suffered by peers; the severed cable is now framed as a memento at company headquarters.

McDonald’s, Vodafone Hit by Azure Credential Theft

A threat actor calling themselves “TheHatman” is selling internal Azure/Entra directory dumps from at least nine Fortune 500-level firms on dark web forums, including McDonald’s (1.7M records), TCS (~800,000), Vodafone (~425,000), and HCL Technologies (~250,000). Hudson Rock researchers say the data — full names, corporate emails, job titles, manager chains, and even Global Administrator account listings — appears highly credible and matches standard Azure export templates.

The exact intrusion vector is unconfirmed, but Hudson Rock linked infostealer-compromised credentials to most affected companies, including a device with dozens of corporate credentials and a direct Kyndryl Azure AD account. Because only large multinationals appear as victims, researchers suspect targeted credential abuse rather than a platform-wide flaw, and warn the exposed org-chart data is ideal for spear-phishing and BEC operations.

Public Exploit Code Released for Microsoft SCCM Vulnerability

Researcher Omri Baso published full proof-of-concept exploit code for CVE-2026-47301, a critical SCCM flaw that chains broken access control, CAB path traversal, arbitrary file write, certificate-bypass, and DLL hijacking to let a low-privileged domain user gain SYSTEM execution on a Primary Site Server. The exploit abuses the SMS_EXECUTIVE service by proxying a malicious DLL through adsource.dll.

The PoC also enables and modifies the built-in RID 500 Administrator account, giving defenders a concrete indicator to hunt for. With working exploit code now public, organizations that haven’t patched face urgent risk since SCCM servers control software deployment and endpoint management across entire Windows fleets — defenders should audit AD permissions on the System Management container and monitor for unusual CAB uploads.

VMware vCenter Attackers Drop JSP Webshell

Attackers are exploiting CVE-2026-59310, a critical vCenter Syslog Server path traversal bug, to gain root code execution without authentication. QUIRSO GmbH mapped 361 affected IPs across 47 countries and assessed with moderate confidence that a Chinese-speaking actor is behind the campaign, which moved from disclosure to mass exploitation within days.

The intrusion chain used malformed cron files for persistence, deployed a JSP web shell and passwordless admin accounts, and ultimately pushed a Babuk-derived ransomware encryptor onto ESXi hosts via the datastore browser, partially encrypting large VMDK files and disabling VMware’s HA agent. Defenders should patch immediately, remove public exposure of management interfaces, and hunt for the extensive IOC set including cron artifacts and reverse-SSH tunneling tools.

Hackers Steal Microsoft 365 Logged-In Sessions

ANY.RUN researchers uncovered Mirage2FA, a Phishing-as-a-Service platform attributed to a group called LinX Coders, that lets users complete legitimate MFA logins before covertly hijacking the resulting session cookie via an Adversary-in-the-Middle proxy. The campaign generated thousands of compromise events touching 9,426 targeted accounts and 3,518 email domains, with the US accounting for 63.7% of victims and technology as the top-targeted sector.

Delivery relies on obfuscated HTML, XHTML, and SVG attachments disguised as HR notices, routed sometimes through Amazon SES. Because the stolen session tokens remain valid regardless of password resets, security teams must invalidate active sessions, enforce FIDO2 keys, and enable Continuous Access Evaluation in Entra ID to neutralize the threat.

GitHub Outage Disrupts Developers Worldwide

GitHub suffered a major worldwide outage beginning around 13:40 UTC on August 17, 2026, degrading Pull Requests, Issues, Actions, Webhooks, and Copilot, with roughly 20% error rates across general traffic and a 50% failure rate for archive and raw content downloads. Enterprise SAML/OIDC authentication and SCIM provisioning were also impacted, threatening SSO access for organizations relying on GitHub Enterprise Cloud.

Microsoft confirmed the worldwide scope, but no root cause was publicly disclosed at the time; the pattern of escalating status updates hinted at infrastructure-level saturation rather than an isolated bug. Developers faced stalled CI/CD pipelines and unreliable issue tracking, with no workaround available besides delaying non-critical operations.

Critical Citrix NetScaler Flaw Bypass Authentication

Cloud Software Group disclosed two critical NetScaler ADC/Gateway flaws: CVE-2026-19490 (CVSS 9.3), an authentication bypass affecting SSL VPN, ICA Proxy, CVPN, RDP Proxy, or AAA vserver configurations; and CVE-2026-19489 (CVSS 8.8), a memory overflow triggered when SIP ALG is enabled within LSN groups, causing denial-of-service. Exploitability of the auth-bypass flaw depends on build version and SAML configuration.

Both flaws affect NetScaler versions before 14.1-73.32 and 13.1-63.21, including FIPS/NDcPP variants, and were responsibly disclosed by a JPMorgan Chase pentester. Given that authentication gateways are the primary control point for remote access, Cloud Software Group is urging immediate upgrades, warning that scanning activity typically follows once technical details circulate.

Mythos 5 Now Available in Claude Security

Anthropic expanded Claude Security (in public beta for Enterprise customers) to use its Claude Mythos 5 model for scanning codebases and returning CWE-classified findings with severity, confidence, and suggested fixes — all requiring human review before patches are applied. The design intentionally limits users to defined defensive outputs rather than open-ended prompting of Mythos 5, reducing the risk of redirecting the model toward offensive use.

Anthropic also launched a $35 million Defender Advantage Fund (0xDAF) providing Claude credits for open-source security remediation, and is expanding its Cyber Verification Program that grants vetted organizations reduced safeguards for authorized security work. The move reflects a broader industry trend of delivering dual-use AI cyber capabilities through controlled, narrowly scoped workflows.

Microsoft SCCM Vulnerability Chained to RCE

XM Cyber disclosed the full attack chain behind CVE-2026-47301 (patched July 14, 2026) affecting SCCM’s AdminService REST API, where a chunked CAB upload endpoint skipped the authorization check applied to standard uploads, letting any authenticated domain user submit malicious archives. Combined with weak signature validation and a “CabSlip” path traversal flaw, attackers could write files outside the intended extraction directory.

The chain culminates in DLL hijacking of the SYSTEM-privileged SMS Executive service via an unsigned adsource.dll, achieving SYSTEM-level code execution from a standard domain account. While Microsoft’s July patch blocked the primary path, users with the Operations Administrator role can still reach the downstream exploit, with a full fix expected in ConfigMgr 2609 in October 2026.

Windows Defender Update for 0-day Vulnerability

A cluster of Security Intelligence updates released August 18, 2026 caused Microsoft Defender’s Quick, Full, and Offline scans to crash, with MsMpEng.exe failing inside mpengine.dll via an access violation. Community researchers speculated the bad rollout may have been a rushed attempt to blunt ShieldBreak, a Defender local privilege-escalation zero-day published just after Patch Tuesday that can raise a low-privileged user to SYSTEM.

Microsoft has not confirmed any connection between the crash and ShieldBreak, and it hadn’t published a formal bulletin at time of writing. Security Intelligence Update 1.457.236.0 (and later 1.457.238.0) restored scanning for many machines, though testing showed the fix wasn’t universal, leaving some users to rely on second-opinion antivirus tools in the interim.

Critical Microsoft Copilot CoSnitch Vulnerability

Varonis Threat Labs disclosed CoSnitch (CVE-2026-24301), a Copilot Personal flaw patched August 18, 2026, that chained an undocumented URL parameter, Copilot’s built-in URL-fetching, and its persistent memory feature to silently exfiltrate data from linked accounts like Gmail and Google Drive after a single click — with stolen data disguised as routine outbound browsing traffic. A poisoned-memory variant could persist even through password resets and device re-enrollment.

Notably, Varonis discovered the exploit through “meta-hacking” — repeatedly questioning Copilot about why the attack “wasn’t possible” until its own explanations revealed the exact undocumented parameter needed. Microsoft found no evidence of pre-patch exploitation, but the case underscores growing risk as AI assistants gain deep access to enterprise data, prompting Varonis to urge treating Copilot like a privileged insider requiring access audits.

Microsoft to Make Passkeys Default in Entra ID

Starting September 1, 2026, Microsoft will automatically enroll Entra ID users currently on SMS/voice MFA into passkey registration prompts, moving away from methods vulnerable to phishing, SIM swapping, and number porting. Microsoft will fully retire native SMS/voice delivery on February 1, 2027, after which affected users face a mandatory, non-optional passkey registration prompt with no opt-out.

Organizations needing to keep SMS/voice must migrate to customer-managed telecom providers via the Microsoft Security Store, with provider info available from September 18 and configuration starting October 30, 2026. Admins can use a temporary opt-out via Microsoft Graph’s passkeyDynamicMigration property, but this only delays — not prevents — the eventual enforcement, making early adoption of passkeys, Windows Hello for Business, and FIDO2 keys the recommended path.

Critical Cursor 0-day Vulnerability

Mindgard disclosed CVE-2026-63093 (CVSS 8.7), a binary-planting flaw in the Cursor IDE where a malicious git.exe placed at a repository’s root runs automatically on Windows the moment the project is opened — no prompts, AI involvement, or prior access required. Cursor initially deemed the report out of scope under a shared-responsibility model but quietly fixed it July 13, 2026, a day before public disclosure, without a formal security advisory.

Further research found the same auto-execution flaw applies to hatch.exe triggered by pyproject.toml files, a stealthier variant since build files don’t typically raise red flags. Cursor ships with Workspace Trust disabled by default, meaning simply opening a folder can trigger code execution; this joins other 2026 Cursor flaws like the DuneSlide sandbox escape, highlighting a growing attack surface in AI-powered IDEs.

Defender Driver Can Be Weaponized to Disable EDR

Check Point Research showed that Microsoft Defender’s legitimate, signed BTR.sys remediation driver can be repurposed by attackers with admin privileges to perform kernel-level file and registry operations by reproducing its undocumented, RC4-encrypted transaction protocol. Since BTR.sys loads early in the “Boot Bus Extender” group — before many user-mode security agents initialize — researchers describe a “golden window” during which it could remove security binaries undetected.

Unlike typical Bring-Your-Own-Vulnerable-Driver attacks, this technique uses a built-in Microsoft component with a valid signature, undermining signature-based trust controls. No in-the-wild abuse has been confirmed, but Check Point’s proof-of-concept tool (BTR_CLI) makes proactive detection via Sysmon Event IDs 15 and 6 essential going forward.

Microsoft 365 MFA and Hijack

TrendAI documented a cloud-only BEC campaign where a fake “PTO Request Denied” email led a finance employee through redirects to an AiTM relay that captured a fully authenticated Microsoft 365 session — no malware involved. Investigators spotted impossible-travel logins from Amsterdam and Los Angeles a minute apart, with MFA already satisfied and no conditional-access triggers.

Over roughly 30 days, attackers impersonated a vendor and then a senior AP colleague via look-alike domains to push fraudulent bank-detail changes for around 20 real invoices, while creating malicious inbox rules to hide vendor collection notices. Defenders should watch for impossible-travel alerts alongside mailbox-rule changes and require dual approval plus out-of-band verification before altering vendor payment details.

CISA Warns Medusa Ransomware Steals Data

CISA, FBI, and HHS jointly updated their Medusa ransomware advisory (AA25-071A), confirming the RaaS operation has compromised over 500 critical infrastructure organizations across healthcare, education, legal, and manufacturing. Affiliates gain access via Initial Access Brokers paying up to $1 million for credentials, or by exploiting known flaws including ScreenConnect (CVE-2024-1709), Fortinet FortiClient EMS, and a new BeyondTrust RCE (CVE-2026-1731), often within 24 hours of disclosure.

Medusa relies on living-off-the-land techniques, vulnerable driver abuse to kill EDR, Mimikatz/CrackMapExec/Rclone for credential harvesting and exfiltration, before deploying its gaze.exe payload that deletes shadow copies and encrypts files with AES-256 under the .medusa extension. Ransom demands reach $15 million with victims given 48 hours to negotiate via Tor or Tox, and agencies urge immediate patching, network segmentation, and phishing-resistant MFA.

12 KB Windows Backdoor Hides C2 Domain

Gen Digital analysts found a minimalist 12,288-byte Windows backdoor disguised as Realtek audio software that hides its C2 domain by counting trailing whitespace characters on lines of a fake desktop.ini file — a technique simple enough to evade string-based and entropy-based scanning. The implant, found on a single Windows 7 SP1 workstation, communicates via an ICMP ping followed by HTTP POST commands.

Persistence relies on a WMI event subscription triggered at a specific clock time rather than at boot, making it dormant until the scheduled hour. Though its C2 infrastructure expired back in 2021, researchers say the case demonstrates that low-prevalence, structurally unusual malware deserves scrutiny beyond hash and domain matching, particularly around WMI filters and whitespace-padded configuration files.

New “Zombie Card” Flaw Lets Expired Visa Cards Make Purchases

University of Massachusetts Amherst researchers presented “Zombie Card” at USENIX Security, an NFC relay attack using two smartphones that revives expired Visa contactless cards for real purchases. The flaw exists because card expiration is enforced only as a terminal-side policy check reading a plaintext field, while Visa’s EMV Kernel 3 excludes that expiry data from cryptographically signed records — unlike Mastercard, Amex, and Discover kernels, which detected tampering and blocked the attack.

Testing across five major US banks in real retail environments found inconsistent issuer-side protection: one bank approved zombie transactions up to $500 because it validated only account and cryptogram data, not the specific card instance. Researchers disclosed findings to Visa in May 2025, with Visa’s red team still reproducing the attack and no confirmed fix deployed, leaving cardholders reliant on physically destroying expired cards.

MessiahGPT Fueling Ransomware and Phishing Attacks

Trellix researchers identified MessiahGPT, an uncensored criminal AI service marketed on BreachForums for generating ransomware, phishing kits, stealers, crypters, and rootkits, claiming to use a 128-expert Mixture-of-Experts architecture trained on unrestricted and dark-web data. The service, tied to messiahgpt[.]de and a Telegram community, offers free trial queries followed by crypto-only subscriptions starting around $8/month, lowering the barrier for inexperienced attackers.

While Trellix could not independently verify the operators’ technical claims, the platform is confirmed active and promoted in criminal circles. The bigger risk is scale: AI-generated phishing and malware variants can bypass static filters, so defenders are urged to focus on behavioral detection — sender reputation, credential-access attempts, and encryption activity — rather than trying to fingerprint “AI-generated” content specifically.

Shell Investigating Data Breach Following Cl0p Ransomware Group Claim

Energy giant Shell activated incident response after the Cl0p extortion group listed it on a dark web leak site, claiming theft of roughly 89 GB of data including engineering drawings, facility photographs, and testing reports. Shell confirmed it is investigating with third-party forensics firms but has not confirmed operational disruption to refineries or drilling operations.

Cl0p, also tracked as TA505/FIN11, is known for pure extortion via mass-exploitation of managed file transfer platforms like MOVEit and Accellion FTA rather than encryption, complicating detection since systems continue operating normally while data remains compromised. Critical infrastructure organizations are urged to audit internet-facing management appliances and enforce MFA across all administrative services.

Hackers Hijack Thousands of WordPress Sites

Check Point Research uncovered StopAndProtect, a campaign that has turned nearly 2,000 hacked WordPress sites into a resilient C2 network, with over 6,000 unique victim IPs found concentrated in the US, Russia, and India. Infection begins with fake CAPTCHA prompts that trick visitors into pasting malicious PowerShell into their terminal, launching a multi-stage .NET-loader chain that deploys ransomware, credential stealers, screen lockers, and a USB-spreading worm.

Operational security failures by the threat actors exposed internal logs and even a personal machine infection, revealing a custom VB6 tool used to manage the hijacked WordPress botnet. One compromised site had gone unpatched since 2021 with nearly 40 vulnerabilities, underscoring how neglected CMS installations are being weaponized as full C2 infrastructure rather than simple phishing redirectors.

The post Weekly Cyber Security Newsletter Bulletin – Entra ID RCE, Claude Code Ransomware, T-Mobile Cable, Azure Credential Theft +20 Stories appeared first on Cyber Security News.