Critical CosmosEscape Flaw Exposed Azure Cosmos DB Primary Keys

🚨 ACTIVE EXPLOITATION

• Critical CosmosEscape Flaw Exposed Azure Cosmos DB Primary Keys and Data Access CVE-2026-66803
  A critical vulnerability in Azure Cosmos DB exposed primary keys, allowing full database access.

  • Applies to Azure Cosmos DB service used by Microsoft products like Entra ID, Teams, and Copilot
  • Vulnerability named CosmosEscape exposed a platform-wide master key granting full read/write access
  • Attack exploited Gremlin API sandbox bypass via .NET reflection to execute arbitrary code
  • Master key allowed enumeration of all Cosmos DB accounts and retrieval of any primary key
  • Issue affected private, network-isolated accounts and Microsoft’s own databases
    📎 Coverage: securityweek.com · 👁 via SecurityWeek

• OpenAI and Hugging Face Breached by Autonomous AI Agent in Linked Attack
  An autonomous AI agent executed a multi-stage cyberattack breaching OpenAI and Hugging Face systems.

  • Applies to OpenAI models including GPT-5.6 Sol and unreleased versions, plus Hugging Face production infrastructure
  • Attack exploited zero-day flaws in internal proxies and data pipelines to run unauthorized code and steal credentials
  • AI agent automated thousands of rapid attempts, bypassing traditional defenses relying on human speed and skill limits
  • Initial access via malicious dataset and proxy vulnerabilities led to privilege escalation and lateral movement
  • Detection occurred post-breach; sandbox isolation and single-step approval checks proved insufficient
    📎 Coverage: cyberscoop.com · 👁 via CyberScoop

🔓 CVEs & KEV

• CVE-2026-10079 — CVSS 8.5 — Stackrox: stackrox: deploy-time policy enforcement and visibility bypass via ...
• CVE-2026-65313 — CVSS 8.1 — Use of hard-coded VNC credentials in the engineering-workstation provisioning...
• CVE-2026-11770 — CVSS 7.5 — 389-ds-base: 389-ds-base: pre-auth ldap filter injection in cleanallruv statu...
• CVE-2026-15722 — CVSS 7.5 — 389-ds-base: 389-ds-base: pre-authentication stack buffer overflow in get_ruv...
• CVE-2026-65310 — CVSS 7.5 — Missing authentication and permissive CORS policyANDRITZ HIPASE-250 (formerly...
• CVE-2026-18436 — CVSS 5.3 — MailerPress <= 1.5.0 - Missing Authorization to Unauthenticated Arbitrary Mod...
• CVE-2026-18437 — CVSS 5.3 — MailPress <= 1.5.0 - Missing Authorization to Unauthenticated Contact Updates...
• CVE-2026-65311 — CVSS 5.3 — Missing authentication for logging-configuration endpointThe HTTP server comp...

🕵️ RESEARCH & DEEP DIVES

• XCSSET v40 macOS malware targets developers via infected Xcode projects
  XCSSET v40 malware infects macOS developers through compromised Xcode projects.

  • Targets macOS developers using Apple's Xcode IDE and open-source GitHub projects
  • Spreads via supply chain attacks embedding malicious loader scripts in Xcode project files
  • Employs polymorphic payloads, fileless persistence, and dynamic in-memory execution for stealth
  • Features 17 modules including Chrome hijacking via Chrome DevTools Protocol and Telegram trojan
  • Infection activates when developers build infected projects locally, enabling credential theft and browser hijacking
    📎 Coverage: unit42.paloaltonetworks.com · 👁 via Palo Alto Unit 42

• DeepSeek-Powered Hermes Agent Conducts Autonomous Cyberattacks on Exposed Servers
  A Chinese-speaking threat actor used an AI-driven agent to autonomously attack exposed internet-facing servers.

  • Targets included exposed Langflow, n8n automation systems, Citrix NetScaler, Marimo notebooks, Apache Tomcat, and VPN endpoints
  • The AI agent autonomously scanned, researched, and launched exploit attempts without direct human input
  • Exploitation attempts focused on vulnerabilities in Langflow and n8n, with some manual attacks successfully stealing data from Citrix NetScaler
  • The Hermes Agent used DeepSeek for reasoning, Telegram for command control, and customized red-team skills to bypass restrictions
  • Researchers gained insight after the attacker accidentally exposed their working environment via an internet-accessible file server
    📎 Coverage: cybersecuritynews.com · 👁 via Cyber Security News

• Critical SolarWinds Web Help Desk SAML Login Bypass Patched in 2026.2.1 CVE-2026-28323
  SolarWinds patched a critical SAML authentication bypass in Web Help Desk.

  • Applies to SolarWinds Web Help Desk using SAML 2.0 single sign-on
  • Vulnerability CVE-2026-28323 allows bypassing SAML authentication
  • Exploitation grants unauthorized access to help desk data and internal info
  • Fixed in version 2026.2.1 released July 30, 2026 with CVSS score 9.8
  • Update also addresses denial-of-service and multiple third-party vulnerabilities
    📎 Coverage: cybersecuritynews.com · 👁 via Cyber Security News

• Coordinated cyberattack targets 30+ Minnesota water utilities via OT vulnerabilities
  A coordinated cyberattack targeted over 30 Minnesota water utilities exploiting operational technology weaknesses.

  • Applies to 30+ small Minnesota community water utilities using operational technology
  • Attack exploited vulnerabilities in programmable logic controllers (PLCs), possibly Rockwell Automation MicroLogix 1400
  • Attack delivered via cellular-connected equipment and targeted industrial control systems
  • Incident caused limited disruption with quick recovery; no water safety impact reported
  • Federal agencies link attacks to Iranian-affiliated threat actors targeting US water infrastructure
    📎 Coverage: csoonline.com · 👁 via @metacurity@infosec.exchange

⚠️ ADDITIONAL ALERTS

• New SSH Bot Profiles Linux Systems Before Deploying Cryptocurrency Miners
  A new SSH bot quietly profiles Linux hardware before deploying cryptomining malware.

  • Targets Linux servers accessed via weak root SSH credentials
  • Performs hardware reconnaissance: CPU, GPU, RAM, OS, uptime, and login history
  • Uses a Go-based SSH client and authenticates with weak passwords like root/123123
  • No immediate payload; bot grades system for mining suitability before later attack
  • Indicators include IP 91.92.40.13, HASSH fingerprint 2ec37a7cc8daf20b10e1ad6221061ca5
    📎 Coverage: cybersecuritynews.com · 👁 via Cyber Security News

• Google AI Finds 13-Year-Old High-Severity Chrome Vulnerability Amid Record Bug Fixes
  Google AI discovered and helped patch a 13-year-old critical Chrome vulnerability.

  • Applies to Google Chrome browser, patched in Chrome 145 and later versions
  • Vulnerability was a sandbox escape allowing compromised renderer to read local files
  • Exploitable via crafted HTML pages due to insufficient data validation in Navigation
  • Discovered using Google's AI agent harness leveraging Gemini and LLMs trained on Chrome's codebase
  • AI-driven detection contributed to over 1,800 Chrome security fixes in 2026, including 370 in latest release
    📎 Coverage: securityweek.com · 👁 via SecurityWeek

• ShutterGap Exposure Lets Attackers Copy Millions of AWS Resources Between Scans
  Temporary public AWS resources can be copied by attackers before cloud security scans detect them.

  • Applies to AWS customers using RDS snapshots, DocumentDB snapshots, AMIs, and SSM documents
  • Vulnerable resources are briefly made public due to customer misconfigurations during creation or modification
  • Attackers monitor and copy exposed snapshots within minutes before CSPM and CNAPP tools scan again
  • Exposure windows often last less than two minutes, too short for daily security scans to catch
  • Copied snapshots contain sensitive data including AWS account IDs, emails, secrets, and financial info
    📎 Coverage: cybersecuritynews.com · 👁 via Cyber Security News

📰 UNDER-REPORTED

• dc-public-schools-student-data-breach-2026 — @metacurity@infosec.exchange