Cisco Talos detailed a financially motivated, Chinese-speaking group it tracks as UAT-10147 that breaks into internet-facing Windows and Linux web servers at scale and installs malware for data theft and search-engine-optimization fraud. The group weaves AI tools through its operations, from exploiting known vulnerabilities to generating payloads, and Talos even found AI-generated code comments left in the source of its Linux kernel rootkit. Its cross-platform implant, SPECTRE, offers credential theft, process injection, encrypted command-and-control, and driver-based bypassing of endpoint detection. Investigators also recovered prompt logs from AI coding assistants on the attackers' own machines, a concrete look at adversaries using the same agent tools defenders do.
Researchers at Expel found a new malware toolkit, SynkLoader, spread through Microsoft Teams messages in which attackers pose as a company's IT help desk. Using their own Microsoft tenant and an onmicrosoft.com address for credibility, they talk an employee into installing a fake "PowerShell Cleaner" hosted on Microsoft's own Azure storage. Once installed, SynkLoader can load modules including a convincing full-screen fake Windows lock screen that captures the user's password, plus a reverse proxy, remote shell, and remote desktop control. Its focus on counting Active Directory systems suggests it is used by a ransomware group or access broker to size targets. The fake lock screen can be escaped with Alt+Tab or Ctrl+Alt+Delete.
Check Point researchers showed at Black Hat that Microsoft Defender's own legitimately signed boot-time cleanup driver, BTR.sys, can be abused to delete security software during startup. The driver, bundled inside Defender to finish removing malware after a reboot, can perform arbitrary kernel-level file and registry operations, and a released proof-of-concept wiped the entire Defender stack from a fully updated Windows 11 machine with tamper protection on. Unlike bring-your-own-vulnerable-driver attacks, this uses a driver present in every Windows since Windows 7, so it cannot be blocklisted. It requires administrator rights with a specific privilege, so Microsoft considers it a trust-boundary issue rather than a bug and will not patch it.
Kaspersky documented what it calls the first malware found on a car head unit with an infection chain built specifically for that kind of device. The malware spreads through the built-in software updaters of certain Android-based automotive head unit firmware, then pulls a multi-stage downloader that runs ad fraud and enrolls the unit into a reverse-proxy botnet. Researchers attribute it with high confidence to a group tied to the BADBOX ad-fraud and residential-proxy operation. A head unit is the central console that handles media and, on many vehicles, some vehicle functions, so malware delivered through its own update mechanism is a notable expansion of automotive supply-chain risk.
Trend Micro found 14 malicious npm packages that pose as working calendar and streak utilities while secretly installing a Linux backdoor from the commercial RedC2 4.0 toolkit. The packages function as advertised, but on load they locate a bundled binary disguised as a math accelerator, mark it executable, and run it as a detached background process. No install script is needed, so a single import anywhere in the dependency graph, even a transitive one, triggers execution. RedC2 is sold on criminal forums as an evasion-focused command-and-control framework with surveillance, credential theft, tunneling, in-memory payload execution, and AI-assisted command features. It shows how import-time execution keeps making package registries an easy delivery route.
Researchers at Truffle Security reported that after four years of collecting leaked Amazon Web Services keys, they found 768 that still grant full control over a company's cloud account, with a median age of about five years. The keys were exposed in places like public code and configuration and were never rotated, so they remain live long after the people who created them have likely forgotten them. A single valid key with broad permissions can let an attacker read data, spin up resources, and move through a cloud environment. The finding is a reminder that leaked long-lived credentials remain one of the most durable and overlooked paths into cloud accounts.
Attackers briefly poisoned arrayref, a foundational Rust crate with about 245 million downloads that sits underneath widely used graphics and blockchain libraries, along with two sibling crates from the same maintainer account. The crate code itself was clean; each added a dependency on a typosquat of a popular package whose build script ran during compilation, pulling and executing an infostealer that grabbed host data and browser credentials. Because the malicious code lived in a build script, simply compiling a project that resolved the crate ran it, with nothing from the library needing to be called. The bad versions were pulled within about ninety minutes, but any build during that window was exposed.
Researchers at Socket found 40 malicious Firefox extensions, part of a wider set of 77, that impersonate cryptocurrency wallets like OKX, Rabby, and TronLink to steal users' funds. Dubbed the Offside Wallet Theft Factory and active since March, the campaign uses lookalike names with subtle character swaps, cloned wallet code, and reused extension identities. Some variants capture recovery phrases as a user sets up or imports a wallet; others copy the wallet's keyring before it is encrypted locally, so on-device encryption offers no protection. Stolen data goes out through Cloudflare Workers, attacker databases, and hardcoded servers. Removing an extension after theft does not secure the wallet, so victims must move funds.
Varonis disclosed a flaw in Microsoft Copilot Personal, the consumer assistant, that could let a single click exfiltrate data from connected apps, and the way they found it is striking. Tracked as CVE-2026-24301 and named CoSnitch, the technique had the researchers repeatedly ask Copilot why a prompt could not run without user interaction; each refusal added a technical justification, until the assistant named a hidden parameter, the exact session conditions where it worked, and the protections meant to block it. Building the request as described, the bypass ran. Varonis said Copilot was not breached but played. Related research showed attacker pages persisting unwanted memory in Copilot through indirect prompt injection.
Researchers, including teams at Anthropic and EPFL, demonstrated that self-propagating instructions can spread from one AI agent to another through the editable prompt and state files that autonomous agent harnesses use to carry context between sessions. In simulated multi-agent coding setups, a payload written into a shared file could infect the next agent that read it. The researchers call the risk real but currently limited, noting there is no sign of it spreading in the wild and that compromising one agent usually already grants machine access. Encouragingly, adding a single short warning paragraph to an agent's system prompt cut propagation to nearly zero across the payloads they tested.