📰 The CyberSignal Daily — August 22, 2026

Microsoft confirmed that a maximum-severity flaw in Entra ID, the identity service standing in front of Microsoft 365 and Azure, was exploited in the wild before the company quietly fixed it on its own infrastructure. Cisco Talos published two reports tying a Chinese-speaking crime group to agentic AI running inside live intrusions, alongside a cross-platform implant called SPECTRE. The Rust Project deleted three poisoned crates.io packages after a compromised maintainer account slipped in a typosquatted dependency whose build script executed malware at compile time. watchTowr's finding that GitLab CVE-2026-19478 is under active exploitation was independently confirmed. And Check Point showed that Defender's own signed boot-time driver can be pointed at security software instead of malware. The thread running through all of it is below.

Identity & Cloud Security

Microsoft confirmed that CVE-2026-69836, a remote code execution flaw in Entra ID carrying the maximum CVSS score of 10.0, was exploited before it was publicly disclosed. The bug is a deserialization of untrusted data issue: a backend endpoint processes crafted serialized objects without validating them, letting an unauthenticated attacker execute code with no user interaction. Because Entra ID is a fully managed cloud service, Microsoft deployed the fix across its own infrastructure. There is no KB, no update package, and nothing for customers to install.

That is exactly why this one is easy to file under "handled" and shouldn't be. The vulnerable component sits in the identity plane that brokers access to Microsoft 365, Azure, and connected third-party apps, and Microsoft has published no exploitation timeline, no scope, and no indicators. Treat the window before the fix as potentially hostile: pull Entra sign-in and audit logs for the period, hunt for unfamiliar app registrations, service principals, and consent grants, and rotate credentials for any privileged identity you cannot fully account for.

→ Read the full story

Threat Intelligence

Cisco Talos published two linked reports on UAT-10147, a Chinese-speaking cybercrime group that targets a wide range of vulnerable web servers. The newsworthy part is not a new piece of malware. It is how the group runs its intrusions: Talos says UAT-10147 has folded agentic AI into its post-compromise operations, using it to drive activity after the initial foothold rather than to generate a payload. The companion report covers SPECTRE, a cross-platform implant spanning both Linux and Windows, and documents its capabilities and indicators of compromise.

The practical value here is the IOC set, which turns a trend story into hunting work you can actually schedule. Start with the exposed web servers UAT-10147 favors, then run the SPECTRE indicators across both Linux and Windows estates rather than just the Windows side where implant hunting usually stops. Two things to hold loosely: whether SPECTRE is sold to other crews or held exclusively by this group is not established in what has been published. Treat any indicator match as a live intrusion, assume a foothold, and open incident response rather than logging it as a curiosity.

→ Read the full story

Supply Chain

The Rust Project deleted malicious releases of arrayref, internment, and append-only-vec from crates.io. A compromised maintainer account had added a typosquatted proc-macro1 dependency whose build script executed malware at compile time, meaning a developer did not have to run the resulting binary to be compromised. A cargo build was enough. The blast radius is wide: arrayref has roughly 245 million downloads and 403 direct dependents, sitting underneath projects including winit, egui, iced, and blake3, plus crates tied to the Solana and Ethereum ecosystems.

Build-time execution is the detail that should change your response speed. Most dependency-hygiene advice assumes the malicious code runs when the application runs, which leaves time to catch it in review or testing. A build script fires on the developer's laptop and in CI, before anyone inspects anything. Check whether any of the three crates appeared in your Cargo.lock during the affected window, and if so, treat those build hosts as credential-compromised: rotate developer tokens, cloud keys, registry credentials, and CI secrets rather than waiting for evidence of use.

→ Read the full story

Vulnerabilities

The Hacker News independently confirmed that GitLab CVE-2026-19478 is under active exploitation within days of disclosure, corroborating watchTowr's original finding and widening the reporting on it. The flaw is a code injection issue rated CVSS 9.4 that an unauthenticated attacker can reach through GraphQL, allowing modification or deletion of publicly accessible projects and rewriting of their data on unpatched self-managed instances. Fixed builds shipped August 17.

Independent confirmation matters more than it sounds. A single vendor's exploitation claim often stalls patch approval in change-management meetings; a second source removes that argument. If you run self-managed GitLab and have not moved to a fixed build, this is the week's clearest deadline. GitLab.com is not affected, so the exposure is entirely on self-hosted instances, which are also the ones most likely to be running behind a maintenance window nobody has claimed.

→ Read the full story

Endpoint Security

Check Point Research disclosed a technique that abuses BTR.sys, Microsoft Defender's legitimately signed Boot Time Removal Tool driver, to perform arbitrary kernel-level file and registry operations. No vulnerability is exploited and no outside driver is imported. The driver is already on the machine, already signed by Microsoft, and already trusted to delete stubborn malware before Windows finishes loading. Point it at different targets and it will delete security software instead. Check Point reports the technique works from Windows 7 through Windows 11 25H2.

This is a bring-your-own-vulnerable-driver attack without the bring-your-own part, which is what breaks the usual defense. Driver blocklists, vulnerable-driver rules, and signature checks all assume the hostile component arrived from somewhere else. Here it ships with the operating system. Shift detection toward behavior: watch for unexpected invocations of the boot-time removal path, monitor tampering with Defender's own service and registry configuration, and confirm that EDR tamper protection and boot-integrity telemetry are actually enabled rather than assumed.

→ Read the full story

Yesterday's stories look unrelated until you ask one question of each: where did the attacker's capability come from? In every case the answer is from inside the trusted set. Entra ID is not something an attacker had to reach past; it is the thing that decides who gets in. The Rust payload arrived through cargo, invited, and ran as part of a normal build. GitLab's GraphQL API is a documented feature, not a backdoor. BTR.sys is signed by Microsoft and shipped with Windows. And UAT-10147's agentic AI is not malware at all; it is an operator, working the intrusion after the door is already open.

That pattern is a problem for how most detection is built. A blocklist asks whether a driver is known-bad. A signature check asks whether a binary is signed. A dependency scanner asks whether a package is on a list of known-malicious names. All three questions assume the hostile thing is foreign, and all three return a clean answer when the hostile use runs through legitimate machinery. A build script executing at compile time is not an anomaly. It is how Rust works.

The move is to stop asking whether an artifact is legitimate and start asking whether its use makes sense in context. Defender's boot-time removal driver is legitimate; it firing on an endpoint that had no malware to remove is not. A cargo dependency is legitimate; a new transitive dependency appearing in a patch-level bump is not. A GraphQL mutation is legitimate; an unauthenticated one rewriting a public project is not. None of that needs new tooling. It needs provenance and baselines, which are harder to buy and easier to postpone.

One concrete consequence this week: the Entra ID fix is finished and you had no part in it, which means your only remaining task is the one nobody will assign you. Go read the logs from before the patch.

All four of yesterday's CyberSignal articles are in the Top Stories above, so these are the strongest pieces from the day before.

Stay sharp. Stay ahead.

Till next time,

The CyberSignal Team

Our News Site | Get the Weekly Briefing