A Max-Severity Entra ID Flaw, Already Exploited (08/21/2026)

Two of today’s stories are about credentials that outlived everyone who remembered creating them. Truffle Security spent four years collecting leaked AWS keys and found 768 that still grant full control of a company’s account, with a median age around five years. The other is the identity system itself: Microsoft disclosed a maximum severity flaw in Entra ID that was already being used against customers before anyone outside knew it existed.

Top 5 Critical Security Alerts

1. A Maximum Severity Entra ID Flaw Was Exploited Before Disclosure

CVE-2026-69836 is a deserialization of untrusted data flaw in Microsoft Entra ID that let an unauthorized attacker execute code over a network, with no privileges required and low attack complexity. Microsoft says it has already been exploited in attacks, that exploit code is not yet public, and that the fix is entirely service-side with no action for customers. The company published the advisory on August 21 and described the CVE’s purpose as providing further transparency, without detailing scope or victims. BleepingComputer

Operator Note: No action for customers is accurate and it is not the same as no consequence for customers. Entra ID is the identity control plane for most Microsoft estates, so a period of unauthenticated code execution inside it is a period where every assumption about who authenticated is worth less than you would like. Nobody can tell you whether your tenant was in scope, because Microsoft has not said. What you can do is check sign-in logs and privileged role assignments over the past several weeks for anything you cannot account for.

2. Researchers Found 768 Leaked AWS Keys With Full Account Control

Truffle Security spent four years collecting exposed AWS secrets from code repositories, Git history, datasets, Docker images, registries, and continuous integration logs. From 431,875 secrets they got 64,024 unique keys, and of the subset they could fully verify, 768 gave complete control of a company’s AWS account, including 526 root keys and 242 carrying the AdministratorAccess policy. Hugging Face alone accounted for 8,482 unique exposures. As of August 10, 88% of the tested subset still authenticated. The median key age was about five years, the oldest had been live 17.4 years, and only 13.7% had ever been rotated. BleepingComputer

Operator Note: That 13.7% rotation figure is the whole finding. These are the machine identities we wrote about in what to fix first in identity and access management, and they behave exactly as described: no owner, no expiry, no monitoring, and no one to notice. A key that has been valid for 17 years was created by somebody who has since left, for a system that may no longer exist. Go pull the list of access keys in your own account with their creation dates, sort descending, and start at the top.

3. GitLab Went From Disclosure to Exploitation in Days

We carried CVE-2026-19478 on August 17 with no known exploitation and technical details held until roughly mid-November. watchTowr has now observed in-the-wild exploitation against its honeypot network. Beyond modifying and deleting public projects, attackers can delete entire repositories, forge merge records so a fix appears to have landed when it did not, and ban project maintainers. The 9.4 affects versions 18.2, 19.0, 19.1, and 19.2 before their patch releases. watchTowr’s Jake Knott notes that AI-enabled attackers are compressing the time from disclosure to exploitation. The Hacker News

Operator Note: Forging a merge record is the detail worth carrying to your engineering leadership. An attacker who can make a fix appear to have landed has attacked your evidence rather than your code, and every downstream control that reads the repository as the source of truth inherits the lie. This is the second time this month a supply chain attack has worked by corrupting what the record says rather than what the software does.

4. Cisco Ships Five Separate 10.0s

Cisco patched nine flaws across Crosswork Data Gateway, Network Controller, and Planning, plus Secure Workload. Five carry a full 10.0: SQL injection, missing authentication for a critical function, external control of the file system, improper access control, and improper authentication. All were found in internal testing and none are known to be exploited. A related Secure Firewall ASA and FTD flaw, CVE-2026-20349, has been exploited in the wild. Fixes are Crosswork 7.2.1-SP and Secure Workload 3.10.9.1 and 4.0.4.16. The Hacker News

Operator Note: Five maximum scores in one advisory is unusual and it is also the good version of this news, because Cisco found them internally rather than learning from a victim. The one to move on first is the firewall flaw that is actually being exploited, which is a different product and a different ticket than the five that are making the headline.

5. Defender’s Own Driver Can Delete Your Security Software

Researchers showed Microsoft Defender’s driver can be weaponized to delete security software at boot, before those products have a chance to start. The Hacker News

Operator Note: A signed driver from the security vendor, used to remove security tooling, running before anything is watching. Tamper protection is the control that is supposed to cover this and it is off or unconfigured in a lot of estates. Go confirm yours is on and that somebody gets an alert when a protection component stops.

Additional Security Alerts

Threat Intelligence

• Poisoned Rust crates are stealing developer credentials, with North Korean actors tied to the campaign: The registry changes and the technique does not. The Register
• Android car malware is spreading through built-in updaters for ad fraud and proxy work: The update mechanism in the vehicle is the distribution channel. The Hacker News
• SynkLoader is being pushed through a Microsoft Teams phishing campaign: BleepingComputer

Security Breaches & Incidents

• US Bank says the breach claims relate to a fourth-party incident: Not their vendor, their vendor’s vendor, which is a scoping problem most third-party programs cannot see. The Record
• Canada’s Hospital for Sick Children was attacked again, with employee data involved: The Record
• Apollo confirmed a data breach amid a wave targeting financial firms: TechCrunch

Security Tools & Best Practices

• Homeland Security is telling users of TrueConf to patch it: The Register
• OWASP published a blueprint flagging the top AI skill risks: Useful scoping for anyone letting an assistant take actions rather than answer questions. Dark Reading

The Axe Report is a daily briefing from Grab The Axe. Need help assessing your organization’s security posture? Take our free Human Attack Surface Score assessment.