Seclog - #189

In this week's Seclog, the security landscape is dominated by critical vulnerabilities impacting core applications and cloud infrastructure, alongside a significant trend in leveraging AI for both offensive and defensive security operations. Immediate threats include a severe arbitrary file read and RCE vulnerability in Ruby on Rails (CVE-2026-66066), which was rapidly weaponized post-disclosure. Further highlighting infrastructure risks, "CosmosEscape" granted full read/write access to Azure Cosmos DB, and a 20-year-old vulnerability provided access to thousands of data center bare-metal servers. A new wave of AI-related risks is also evident, from the discovery of over 200,000 live credentials in Hugging Face AI training data to methods for jailbreaking local AI models with JSON and authentication bypasses in Google AI Studio. Conversely, AI is being rapidly adopted for defense, with new tools for LLM-based reverse engineering, AI-native threat intelligence platforms, and automated vulnerability management agents being deployed in production environments like Figma. These developments collectively underscore the escalating sophistication of threats and the increasing reliance on AI to both analyze and combat them across diverse attack surfaces.

📰 SecLinks #

KindaRails2Shell: RCE via MATLAB File - ethiack.com

A critical vulnerability, CVE-2026-66066, in Ruby on Rails allows arbitrary file read and remote code execution as root. This exploit chain leverages a .mat file declared as image/png to bypass security controls and achieve RCE on default Rails applications.

Figma Leverages AI Agents for Vulnerability Management - figma.com

Figma implemented "agents" that provide continuous vulnerability management by guarding code during writing, reviewing pull requests, and auditing a decade-old monorepo. This approach demonstrates a proactive, automated strategy for enforcing security policies throughout the software development lifecycle.

Frontier Lab Agent Intrusion Technical Timeline - huggingface.co

This detailed technical timeline provides an in-depth analysis of the July 2026 Frontier Lab Agent intrusion incident. The report offers critical insights into the Tactics, Techniques, and Procedures (TTPs) utilized by advanced persistent threats targeting AI/ML research and development environments.

20-Year-Old Vulnerability Exploits Data Centers - lavahq.io

Researchers exploited a 20-year-old vulnerability to gain access to thousands of bare-metal servers within data centers globally. This highlights the severe and persistent risk posed by legacy flaws in critical infrastructure, providing deep footholds for attackers.

Jailbreaking Local AI Models via JSON - linkedin.com

A method for jailbreaking local AI models, specifically Apple's "Guided Generation" in macOS 27, is described by leveraging JSON structures. This technique represents a new attack vector for bypassing safety mechanisms and controls in on-device AI applications.

Google AI Studio Authentication Bypass - ndevtk.github.io

A writeup details an authentication bypass vulnerability affecting Google's AI Studio (aistudio.google.com). This critical flaw could allow unauthorized access to AI development environments and associated resources, posing significant security risks.

AI-Native Threat Intelligence Platform Introduced - beyondmemory.io

BeyondMemory launched an AI-native threat intelligence platform, BeyondCTI-27B, that unifies dark web monitoring, stealer log analysis, and attack surface reconnaissance. This platform provides comprehensive, AI-driven insights for proactive security defense and threat hunting.

S3 Vector Security Analysis for LLM RAG - offensai.com

An analysis examines the security model of Amazon S3 Vectors, specifically for LLM Retrieval Augmented Generation (RAG) pipelines. It highlights critical considerations such as access control, input validation, metadata integrity, and audit coverage to prevent data poisoning and other attacks.

Metasploit Framework 6.5 Adds AI Tooling, C2 Profiles - rapid7.com

Metasploit Framework 6.5 is released with Malleable C2 profile support, AI-tooling integration via MCP server, and NTLM relaying upgrades. These enhancements provide advanced capabilities for penetration testers and red teams, including better threat emulation and evasion techniques.

Rails CVE-2026-66066 Attack Details, Forensic Tools - discuss.rubyonrails.org

The Rails security team has released comprehensive details for CVE-2026-66066, including attack vectors and forensic investigation tools. This disclosure was accelerated due to researchers rapidly reverse-engineering and publishing proofs-of-concept for the critical vulnerability.

Reflecting on Frontier Class Vulnerabilities - shubs.io

This article offers a personal reflection on "frontier class vulnerabilities," drawing from early career experiences with challenging source code review engagements. It suggests these vulnerabilities are significant, complex, and represent a continuous evolution of difficult-to-address security issues.

Hugging Face Data Reveals 200K Live Secrets - trufflesecurity.com

Truffle Security scanned 7.6 petabytes of public datasets on Hugging Face, uncovering over 221,000 live credentials, including tokens capable of pushing code into widely installed software. This discovery highlights a critical supply chain risk and the significant exposure of secrets within AI training data.

Apple Screen Sharing Pre-Auth RCE Disclosed - warez.sl0p.foo

A pre-authentication Remote Code Execution (RCE) vulnerability affecting Apple Screen Sharing on macOS versions up to 26.5 has been disclosed. This critical flaw allows attackers to execute arbitrary code on vulnerable systems without requiring prior authentication.

CosmosEscape Grants Full Azure Cosmos DB Access - wiz.io

Wiz Research discovered "CosmosEscape," a critical vulnerability in Azure Cosmos DB that granted full read/write access to every database within the service. Although fully remediated, this incident underscores the severe impact of critical vulnerabilities in major cloud services.

💻 SecGit #

Radare2 Integrates LLM-Based Reversing Capabilities - github.com

The r2ai project integrates LLM-based capabilities into radare2, enhancing reverse engineering workflows for security professionals. This tool leverages artificial intelligence to facilitate more efficient and intelligent analysis of binaries.

Suzaku: Sigma-Based Cloud Forensics Tool - github.com

Suzaku (朱雀) is an open-source tool designed for sigma-based threat hunting and generating fast forensic timelines from cloud logs. This project aims to assist security teams in quickly analyzing and correlating events in cloud environments for incident response and threat detection.

AI-Powered Reverse Engineering Skill Pack - github.com

The "reverse-skill" GitHub project offers an AI-powered router pack for reverse engineering, authorized penetration testing, and security research. It supports various AI coding clients and provides on-demand toolchain bootstrapping and a self-evolving knowledge base for security professionals.