Security Review #327

We deonstrate that most package managers have an integrity problem. Out of the 25 most-popular managers we tested, they either lack default lockfile verification or signature validation, allowing a malicious repository to inject code into legitimate packages.

Building a custom C# PowerShell runspace that bypasses AMSI and ETW while enforcing Full Language Mode. Breaking down each defensive layer and how operators can neutralize them using .NET reflection from a compiled binary.

PaperCut CVE-2026-81578 (Authentication Bypass) and CVE-2026-82078 (Post-Auth RCE) were exploited in-the-wild, and subsequent patch-bypass flaws (CVE-2026-82077) enable a full pre-auth remote code execution chain. The chain lets an unauthenticated attacker bypass authentication, reset the admin password and execute arbitrary OS commands via the Scan-to-Fax functionality.

MyReport is a report generation software often used as a component of web-based applications. We describe a privilege escalation and remote code execution vulnerabilities in MyReport Updater through its update configuration and gRPC service.

W analyze an incident involving EtherRAT, a Node.js-based backdoor, delivered through a malicious Advanced IP Scanner MSI. It was imported into Microsoft Intune and silently installed on four managed endpoints. The malware collected system information, enumerated network sessions, captured the primary display, and queried public Ethereum RPC services for C2 configuration.

We detail a full chain starting from an LG-signed manifest, a SECURE-tier client-key generation, a cmp-client force browsing, and an SSRF to end into a remote command execution.

We provide an overview and IoC for Installer.exe, a signed binary with an EV certificate shows a fake "verify you're human" slider that always fails while it silently downloads and installs a second, differently signed package that plants an "updater". It does whatever a file on a Cloudflare R2 storage bucket tells it to, including downloading and launching new programs.

Valen's new flexible borrow checker implements group borrowing, a relaxed borrow-checking approach that tracks reference paths to detect use-after-free at compile time and permits multiple read-write references to the same object. The technique provides zero-runtime-cost memory safety while supporting patterns such as multiple mutable locals, parameter-inside-object borrowing, and seamless interop with Rust code.

How we pwned a Linux CHECKSUM_PARTIAL trust-boundary bug: a post-validation IPv4 header mutation, a 32-byte host-memory disclosure, and an owned-host VM escape.

fbclid is a URL parameter appended by Meta's Facebook and Instagram that encodes a platform header, named fields and an optional AEM suffix. Decoding enables investigators to identify the exact click time, source platform/app, and correlate multiple clicks to the same browser or advertising campaign.

CVE-2026-103572 is a remote code execution vulnerability in the shared Microsoft.VisualStudio.Copilot.Core.dll that affects Visual Studio and SQL Server Management Studio, where a trailing-dot filename bypasses path validation and allows Copilot to write to its own settings files and execute arbitrary commands.

This article explains Microsoft's Active Directory tier model and how to enforce tiered administrative boundaries - using OU layouts, deny-logon GPOs, Protected Users, time-bound group membership, and privileged access workstations - to prevent lower-tier accounts from accessing or exposing Tier 0 credentials and thereby reduce the risk of credential theft and unauthorized control of the directory.

Lifting Lua to LLVM IR, why dynamic table operations survive optimization, and recovering Luraph-protected scripts with AI-assisted tooling.

An overview of real intrusions run by autonomous AI agents, recovered from the agents' own attack transcripts. Operators set targets and rules; the agents ran the full loop from recon to root. Assessed as at least two threat actors, likely more.

Technical analysis of CVE-2026-105133 and CVE-2026-105134 in AhsayCBS, revealing how authentication bypass and path traversal enable remote code execution.

Chaining an unauthenticated bulk import, an in-band XXE, leaked IMS credentials and Groovy execution into full pre-auth RCE on Adobe Experience Manager Cloud.

We introduce Entra ID tiering model, provide main security recommendations and highlight how mis-labelled or seemingly unprivileged roles (e.g., Agent ID Administrator, Directory Synchronization Accounts) can still grant broad access. Therefore review of role assignments, application owners, and Graph permissions is essential for protecting tenant administration.

In the first part of the series we focus on a NULL pointer dereference vulnerability. We will utilize a real-world Linux kernel vulnerability. We will walk through the entire exploitation process - from setting up the lab environment and triggering the vulnerability to maintaining system stability and, ultimately, bypassing various mitigations.

CVE-2026-92392 is a use-after-free vulnerability in curl HTTP/2 upload handling that can corrupt a dangling easy-handle pointer and enable remote code execution.

We delve into CVE-2026-81983, an Adobe Acrobat integer overflow vulnerability that corrupts memory when a malicious PDF is opened.

By pivoting on a single Criminal IP favicon hash, this investigation uncovers the "Olho de Deus" C2 infrastructure.

CVE-2026-83498 affects Microsoft Defender SmartScreen's Enhanced Phishing Protection enclave (SFAPE.dll), where a repeated-initialization bug allows a fixed-value write that corrupts the enclave's range validator and enables arbitrary read/write primitives leading to full VTL1 code execution and control of the enclave (and potentially the Secure Kernel).

An unescaped separator in Telegram Desktop's single-instance IPC lets one clicked link read arbitrary files off the disk and send them to the attacker, session files included.

A poisoned CRM lead or a malicious insider could make Salesforce Agentforce post phishing links in Slack with no approval and no trace of who sent them.

Analyzing Eufy Security Ecosystem and reverse engineering of its Homebase2 component. We walk through firmware extraction, Wifi credential generation algorithm, rubber-hose treatment to the UART lockdown, and filesystem gardening for debugging purposes.

We found a 0-click attack on Salesforce Agentforce. A prompt injection in a Web-to-Lead form could bypass URL redaction and leak CRM data over DNS.

We explore how a malicious USB drive is built, what kind of code is written onto it, and how attackers rely on human behavior to get a Windows or Linux user to plug it into a computer.

Harnesses are orchestration layers around LLMs that control inputs, tools, prompts, models, state, validation gates, and outputs, enabling structured multi-stage security research pipelines. In this article, we first review a set of publicly available security harnesses, then detail how to desifn and build your own harness.

In this post I will walk through each major milestone that has been taken in my quest to build LLM-generated Mythic agents from prompt to deployment. Starting at the beginning, I’ll show what worked, what didn’t, and where this space is likely going next.

CVE-2026-46242 is a use-after-free in __ep_remove(), affecting Android 15 on Pixel 10 and providing an arbitrary read and write primitive. The write can be chained to overwrite credentials and the SELinux SID, yielding a root shell in the vold domain (uid 0, u:r:vold:s0).