Security briefing from the BlueTeamSec community on infosec.pub for Sunday the 1st of November 2026. A busy one today, with thirty stories ranging from Iranian freezer hacks to North Korean billion-dollar heists. Let's get into it.
Pluto Security report that a cryptominer campaign on Hugging Face is back after takedown — attackers are distributing malicious AI models that exploit the trust remote code flag to run miners on self-hosted inference platforms like Xinference. Over thirty-two thousand downloads across two waves, which suggests people aren't checking what they're pulling in.
LevelBlue have written up a wave of AI-assisted phishing kits — BlackApex, SPY, and others — that abuse Microsoft's legitimate device code authentication to bypass MFA. Business-themed lures like SharePoint and Zoom trick victims into authorising the attacker's session. One for anyone managing Entra ID environments.
Air Security found malware in the wild targeting AI coding agents — attackers poison GitHub search results with fake repositories that trick agents into executing commands from malicious READMEs, deploying a fileless PowerShell backdoor using NetSupport Manager. The attack specifically targets Windows systems and establishes persistence through registry modification.
Hacktron published a postmortem of vulnerabilities they found in GNU Savannah — twenty-six flaws chaining from email injection to remote code execution and ultimately root access via kernel exploits. The attack path could have enabled backdooring of Bash and Coreutils, which would have been a supply chain event affecting millions of systems.
Alpine Linux disclosed attempted exploitation of a critical vulnerability in the Onigmo regex engine used by Ruby — their GitLab instance was targeted on the twentieth and twenty-first of October with crafted CI configurations that crashed workers and possibly achieved remote code execution. This one's a CVSS ten, and it affects any application compiling regex from untrusted input.
The GSA Inspector General found that Controlled Unclassified Information, including personally identifiable information, was improperly exposed in the agency's Google Drive due to technical misconfigurations and inadequate oversight. The Data Loss Prevention tool failed on multiple fronts — incomplete search terms, unused optical character recognition, file size limits, and process failures.
Microsoft report attackers are impersonating help desk staff via Teams to gain initial access, then exploiting Active Directory Certificate Services misconfigurations to escalate privileges and take over domain controllers. The techniques involved are ESC1 and ESC8 — worth flagging if your organisation uses external Teams federation.
A trojanised version of the Claude Code extension for Cursor was published to the marketplace on the twenty-ninth of October and stayed up for eleven hours, potentially reaching five thousand developers. Microsoft's writeup covers the persistent SSH backdoor it installed on both Windows and macOS environments.
The US Department of Justice announced that an Israeli national has been extradited for exploiting the Nomad Bridge smart contract back in August 2022, stealing nearly three million dollars in cryptocurrency. The exploit leveraged non-authenticated transactions to bypass security controls, followed by laundering through multiple conversions.
And another crypto heist — this one significantly larger. Technical analysis from drk.vc describes how attackers drained billions from Binance hot wallets across multiple blockchains, including over a hundred and ninety-seven thousand ether and more than eighteen thousand bitcoin. The attack leveraged a novel exploitation of Ethereum's delegation standard through nonce bracketing to compromise the signing infrastructure rather than stealing private keys directly.
Chainalysis attribute the Binance exploit to TraderTraitor, a North Korean-linked group, with the total theft reaching six point four billion dollars. The attackers used funded wallets to submit valid signed authorisations, enabling code execution on hot wallets across multiple chains. Two point eight billion was successfully frozen through coordinated industry response, but the rest employed novel automated laundering techniques.
Andrew Mohawk has a technical teardown of the PHP2Shell vulnerabilities affecting PHP versions prior to recent patches — memory corruption during multipart form data parsing that can lead to remote code execution. The flaws occur during low-level request parsing, which means basic WAF rules looking for typical payload patterns might not catch it. Patching to secure versions is the priority here.
A researcher has disclosed a local privilege escalation vulnerability in macOS that allows unprivileged users to execute commands as root by exploiting a launchd job and injecting commands into a root-privileged Bash process. Apple fixed it in macOS twenty-seven point two by removing the vulnerable configuration. It's being called the simplest macOS privilege escalation to date.
GovExec report that Iranian-linked actors are suspected of cyberattacks targeting commercial refrigeration systems at roughly six US military commissaries, successfully overriding temperature controls. The attacks exploit known vulnerabilities in internet-accessible industrial control systems managing freezer supervisory systems — similar tactics were used in France against three hundred and fifty establishments.
Microsoft Threat Intelligence report that North Korean threat actors are actively integrating generative and agentic AI technologies into their cyber operations tradecraft. This represents an evolution in how nation-state actors are adopting emerging AI tools to enhance offensive capabilities.
On a similar note, the US Attorney's Office for the Eastern District of Virginia announced that a defence contractor employee has pleaded guilty to facilitating a North Korean IT worker scheme. The employee allowed his identity to be used to obtain remote IT positions at nine US companies, receiving laptops at his home and enabling North Korean workers to remotely access corporate networks — all while holding an active Top Secret clearance.
Qianxin's twenty twenty-six mid-year report identifies key trends including AI-enhanced attacks in phishing and evasion, increased targeting of critical infrastructure and identity providers, and continued evolution in ransomware sophistication. Published on the twenty-eighth of October, it provides a comprehensive overview of the threat landscape for the first half of the year.
Google Threat Intelligence and Mandiant report that threat actors are deploying LLM-powered autonomous offensive frameworks — ARTEX, Hexstrike, and CAIRN — in real-world intrusions as of August. These agentic platforms automate complex attack chains at machine speed, affecting government, technology, manufacturing, and telecoms sectors globally. Adversaries range from financially motivated extortion groups to suspected Chinese-nexus espionage clusters.
Sudorem present a detection framework for identifying AI-aided and agentic attacks that leverage large language models to automate reconnaissance, decision-making, and tool adaptation. The framework proposes six behavioural indicators including dependent permutation and software versioning mismatches to detect AI involvement in intrusions. Particularly relevant for MDR environments facing adversaries using AI to accelerate attack loops.
The ofndb project provides a database of OriginalFileName metadata for over a thousand Windows binaries parsed from nearly six thousand builds, designed to combat masquerading techniques. Threat actors commonly rename legitimate binaries to evade detection, but the OriginalFileName field embedded in headers is difficult to alter. This resource enables defenders to detect renamed living-off-the-land binaries through header inspection rather than relying on file paths.
MANTIS is a new open-source threat intelligence workbench platform for defensive security teams — it provides a centralized hub for SOC analysts, threat intel teams, incident responders, and reverse engineers to consolidate workflows, moving from raw data collection to automated analysis and reporting in a unified environment.
ESET have analysed Amatera, a Windows-based infostealer operating as malware-as-a-service since twenty twenty-four, with prolific activity in Latin America during early twenty twenty-six. It spreads through social engineering including fake CAPTCHA checks and ClickFix techniques, exfiltrating credentials, session cookies, cryptocurrency wallets, and development secrets.
dragon-tales is a new open-source binary analysis library built on LLVM twenty-one, providing disassembly, assembly, and analysis capabilities across six architectures including x86, ARM, and RISC-V. It supports symbolic execution with Z3, LLVM IR lifting, FLIRT signature matching, and an intermediate language for architecture-independent analysis. One for malware analysts and reverse engineers.
AbliteratedOutTheBox is a fork of llama.cpp that bypasses LLM safety alignment at inference time through real-time activation editing, without modifying model weights. The tool uses a two-pass detection and patching system to neutralize refusal mechanisms across a hundred and thirty-one supported architectures, which makes static security audits ineffective.
PngLoader is an open-source steganographic framework that conceals arbitrary binary payloads within PNG files using nibble-to-ASCII encoding and Fibonacci-based least significant bit manipulation. It's designed for educational purposes in security research and red team operations, allowing covert payload delivery that evades traditional static file signature detection. Defenders can monitor for Windows API calls like VirtualAllocEx and WriteProcessMemory.
There's a new fork of Impacket available that implements an LDAP signing bypass capability. Impacket is a Python library providing programmatic access to network protocols commonly used in Windows environments, including SMB, MSRPC, LDAP, and various authentication mechanisms. While intended for security research and penetration testing, it's worth noting this provides easy-to-use implementations of protocols frequently exploited in network attacks.
ScatterSteg is an open-source steganography tool that embeds encrypted data within PNG and JPEG images using content-adaptive algorithms. It employs encryption via libsodium and is designed to resist detection better than legacy tools like steghide, though advanced deep-learning detectors such as SRNet can still identify it. It hasn't undergone independent security audit and lacks deniability features.
A researcher has demonstrated a novel credential harvesting attack that bypasses PDF JavaScript sandboxing by wrapping malicious PDFs in HTML pages with WebSocket-based command and control. The technique uses fake Windows eleven authentication prompts and Cloudflare Tunnels for stealthy communications, circumventing CORS protections and browser security restrictions. Particularly effective against Outlook Mobile App users.
Fortra have written up a sophisticated phishing campaign targeting OpenAI users that employs a live-operator model to steal payment card data and bypass multi-factor authentication in real time. The operation uses compromised WordPress sites on Nexcess infrastructure with rapid rotation, polling-based anti-analysis mechanisms, and dynamic OTP theft capabilities.
RedOps researchers demonstrated a technique to use Azure API Management Consumption tier as a policy-based redirector for Cobalt Strike command-and-control traffic. The method leverages native policy engine capabilities to filter and forward malicious traffic through legitimate Microsoft infrastructure, potentially bypassing reputation-based security controls. This red team technique requires public HTTPS endpoints but allows threat actors to hide infrastructure behind trusted Azure domains.
And that's your briefing for Sunday the 1st of November. Attribution goes to the original article authors — the analysis here was automated. Stay safe out there.