Daily Security Intelligence Briefing — July 30, 2026

1. FakeAgent Campaign Uses Claude.ai Artifact Redirects to Deploy SectopRAT

Source & Article Link: Huntress Threat Research / TLDR InfoSec

Technical Severity / Domain: Application & Endpoint Security / High-Critical (29 Organizations Impacted)

Why You Must Read This: Sponsored Bing search ads for "claude desktop app" redirected victims to a genuine claude.ai user-generated artifact URL. The artifact contained JavaScript that forwarded users through claude.ai.download-app[.]us to download a malicious ClaudeDesktop.exe. The executable was actually JetBrains jcef_helper.exe DLL-sideloading a tampered libcef.dll (VMProtect packed) that pulled a .NET SectopRAT payload via EtherHiding from an Ethereum BSC smart contract. A secondary persistence mechanism dropped signed IBM SPSS binaries (sslconf.exe) alongside a malicious tempdir.dll that executed DirectX shader-based AES-256 decryption to defeat sandbox analysis.

Recommended Immediate Action: Block C2 IP 2.24.131[.]246 and BSC contracts; hunt for ClaudeDesktop.exe writes, scheduled tasks under %APPDATA%\Roaming\Microsoft\EdgeUpdate\Install, and Defender exclusions added around software installers. Treat vendor artifact landing pages as untrusted download sources.

2. 24,000+ Data Center IPMI/BMC Hosts Exposed to Offline Password Cracking

Source & Article Link: Lava Research / TLDR InfoSec

Technical Severity / Domain: Infrastructure & Hardware Security / Critical (CVE-2013-4786)

Why You Must Read This: Internet scanning of UDP port 623 revealed 36,872 exposed Baseboard Management Controller (BMC) interfaces, with 24,650 returning password-derived HMAC-SHA1 hashes during RAKP authentication without requiring successful login. Supermicro factory 10-character passwords (26^10 keyspace) were cracked offline in ~1 hour on an 8-GPU rig, while HPE iLO 8-character hashes were cracked in ~32 seconds. Active exploitation is confirmed, including iLO 4 management consoles defaced with 0.3 BTC ransom demands.

Recommended Immediate Action: Block UDP port 623 at network perimeters, rotate all factory BMC/iLO passwords immediately, disable IPMI 1.5 and Cipher Suite 0/NONE, and restrict management interfaces to an isolated VLAN or VPN.

3. HollowGraph Malware Uses Microsoft 365 Calendar Invites for Covert C2

Source & Article Link: Patrick Duggan Research / Unsupervised Learning

Technical Severity / Domain: Cloud Security & Threat Hunting / High (Graph API C2)

Why You Must Read This: HollowGraph malware infected 12 enterprise systems and established command-and-control entirely inside Microsoft 365 using calendar invites as the payload transport channel. Because the communication relies on legitimate M365 infrastructure and API calls, traditional network perimeter firewalls and IP blocklists cannot detect or block the C2 traffic.

Recommended Immediate Action: Implement Graph API behavioral logging and audit for unusual calendar event creations or automated attendee updates originating from internal service accounts.

1. From Zero Trust to "Zero Custody" for Autonomous AI Agents

Domain: AI & Cloud Security Architecture

Subject / Topic Summary: Security analysis argues that Zero Trust model validation is insufficient for AI agents because identity verification cannot prevent prompt injection or in-context hijacking. The proposed "Zero Custody" architecture dictates that agent reasoning loops must hold no secrets or high-privilege credentials directly; instead, external API brokers and security gateways inject tokens and re-identify data only after verifying policy constraints.

Key Technical & Strategic Takeaways:

AI reasoning environments should be treated as ephemeral, untrusted compute spaces.

Credentials, API keys, and raw sensitive data must remain isolated in external broker gateways.

Essential Direct Quote: "Zero Trust is insufficient for AI agents because verifying identity does not stop attackers from hijacking behavior through malicious language in context. Zero Custody holds that the reasoning loop should hold nothing that matters, so a compromised agent has empty hands."

Source: 2am Security / TLDR IT

2. Model Context Protocol (MCP) 2026-07-28 Revision Released

Domain: AI Infrastructure & Architecture

Subject / Topic Summary: The Model Context Protocol (MCP) received its largest architectural update since launch. The 2026-07-28 revision converts MCP into a completely stateless protocol, enabling remote servers to run on serverless/edge infrastructure and scale horizontally behind standard load balancers.

Key Technical & Strategic Takeaways:

Eliminates stateful session stickiness requirements for remote MCP server deployments.

Establishes formal extension specifications for enterprise authentication and auditing.

Essential Direct Quote: "MCP 2026-07-28 is the largest update to the protocol since launch... MCP is now stateless, so it can be deployed on serverless and edge infrastructure, or scaled horizontally behind any load balancer."

Source: MCP Protocol Announcement / TLDR AI

3. Distributed npm Package Cluster Delivers Cross-Platform RAT to Alibaba Developers

Domain: Supply Chain Security & Malware Analysis

Subject / Topic Summary: Socket discovered an 18-package npm typosquatting cluster impersonating Alibaba's @ali scope (smart-config-manager, cloud-config-fetcher, local-config-parser). The packages split loader logic across dependencies and escaped the Node.js VM sandbox via items.constructor.constructor to drop aone-cli, a cross-platform RAT supporting DingTalk lateral movement, file transfer, and encrypted reverse TCP proxying.

Key Technical & Strategic Takeaways:

Persistence achieved via ~/.zshrc entries, 10-minute macOS Launch Agents, trojanized app.asar files, and Linux /tmp binaries.

Remediation: Treat impacted systems as fully compromised, rotate credentials from clean hardware, and audit for suspicious npm dependencies in internal builds.

Essential Direct Quote: "Socket uncovered a three-month stealth campaign involving 18 npm packages that split loader functions across dependencies... escaping the Node.js vm sandbox via items.constructor.constructor to stage a cross-platform RAT."

Source: Socket Threat Research / TLDR InfoSec

4. Claude Mythos Preview Discovers Cryptographic Flaws in AES & HAWK

Domain: AI Security & Cryptographic Research

Subject / Topic Summary: Anthropic detailed research where Claude Mythos Preview discovered new lattice symmetries to recover HAWK-256 signature keys and accelerated 7-round AES-128 meet-in-the-middle attacks by 200x–800x compared to prior human research.

Key Technical & Strategic Takeaways:

Highlights LLM reasoning capabilities in automated mathematical and cryptographic vulnerability discovery.

Production cryptography (full 10-round AES-128 and HAWK-512/1024) remains secure and unaffected.

Essential Direct Quote: "Claude Mythos Preview model found a new lattice symmetry to fully recover HAWK-256 signature keys and cut the cost of a seven-round AES-128 meet-in-the-middle attack by 200 to 800 times."

Source: Anthropic Research / The Hacker News

5. Microsoft Project Perception Multi-Model Defense Stack

Domain: Security Leadership & AI Operations

Subject / Topic Summary: Microsoft detailed its Project Perception architecture combining specialized models (MAI-Cyber-1-Flash) with GPT-5.4. The multi-agent stack coordinates automated red-team attack path generation, blue-team investigation, and green-team remediation across M365, Azure, and identity providers.

Key Technical & Strategic Takeaways:

Public preview begins August 3, 2026 for MDASH customers.

Demonstrates 96% score on CyberGym benchmark while reducing reasoning costs by ~50%.

Essential Direct Quote: "Microsoft's Project Perception is an AI cybersecurity system that coordinates red, blue, and green team agents to defend against AI-driven attacks."

Source: GeekWire / TLDR IT

Tier 3 — Industry Noise & Awareness Only

• FENGARDE Open-Source SIEM Release: SIEM parsing 16 logs into OCSF with local Ollama LLM triage (GitHub FENGARDE).

• Snowflake Cortex AI Gateway Launch: Centralized model governance and cost management gateway for AI agents (VentureBeat).

• Bank of Baroda Customer Data Leak: Employee email compromise exposed customer ID numbers and loan records (TLDR InfoSec).

• Origin Energy Australia Breach: Data security incident exposed partial credit card and financial records for ~900k users (TLDR InfoSec).

• Phishing-as-a-Service AI Kits (TokenLover & YaksaLover): Turnkey AI kits automating M365 BEC and device code phishing (Eye Security).

• Claude Desktop Screen Recording Skills: Desktop feature converts user screen recordings into reusable agent skills (@ClaudeAI).

• Cloudflare Open Sources pvcli Privacy Proxy: Apache-2.0 CLI tool for debugging Oblivious HTTP protocols (Cloudflare Blog).

• DEF CON Bans Smart Glasses: DEF CON 2026 bans camera-equipped smart glasses over recording concerns (The Register).