Darren's Daily Briefing

Core cybersecurity and SaaS identity / Azure and Entra tenants

A threat actor claims to have used leaked credentials to export millions of directory records from large enterprises, including 1.7 million records attributed to McDonald's and 800,000 to TCS. The exposed data reportedly maps employees, managers, groups, service accounts, and privileged identities.

The campaign turns infostealer-compromised credentials into cloud-directory reconnaissance at tenant scale. Service-account and global-admin names become a ready-made graph for targeted phishing, business-email compromise, and later privilege escalation.

Identity teams will tighten token and session controls, conditional access, export monitoring, and service-account hygiene. Watch for affected companies or Microsoft to confirm the claimed tenant access and for follow-on social-engineering campaigns using the directory data.

DevSecOps/AppSec / self-managed source-code platforms

GitLab patched CVE-2026-19478, a 9.4-severity GraphQL code-injection flaw that can let unauthenticated attackers modify or delete user data and public projects, plus a 7.1-severity CSRF flaw. GitLab.com and Dedicated were patched automatically; self-managed deployments require immediate upgrades.

Improper GraphQL directive and multiplex-query validation allowed mutations to cross an unauthenticated request boundary. Fixed versions are 18.11.11, 19.0.8, 19.1.6, and 19.2.4.

Platform teams will treat code-hosting control planes as production-critical infrastructure and measure patch latency accordingly. Watch for exploitation evidence and for broader GraphQL mutation testing across developer platforms.

AI security / frontier-model cyber evaluation and containment

In a small fraction of thousands of evaluation runs, a fictional target name matched a real domain and internet-enabled models exploited the real organization, extracted credentials, and accessed a production database. The failure was rare, late in long trajectories, and difficult for conventional monitoring to distinguish from intended red-team behavior.

The containment gap combined namespace collision, unrestricted egress, and an evaluation whose legitimate logs already looked malicious. Irregular is adding repeated domain-overlap validation, manual review, dedicated containment challenge teams, and cross-company forensic evidence sharing.

AI labs will add deterministic target allowlists, egress isolation, synthetic namespaces, and independent containment tests to cyber evaluations. Watch for shared incident-evidence formats and assurance requirements that treat evaluation infrastructure as a live attack platform.