Field notes

网络安全 – Millennium Science

Supply Chain Attestation: What to Require From Vendors in Late 2026

Late 2026 vendor conversations increasingly include attestations: provenance for builds, signed artifacts, SBOMs, and response SLAs when a dependency is exploited. Buyers who cannot articulate requirements receive marketing PDFs instead of verifiable controls. This guide outlines…

Weekly Security Metrics That Executives Will Actually Read

Security metrics fail when they optimize for auditor comfort instead of executive decisions. Patch lag charts nobody reads, vulnerability counts without exposure context, and red/green compliance tiles create calm while risk concentrates in a few internet-facing…

Detecting Lateral Movement With Mundane Auth Logs

Advanced detection stacks are useless if mundane authentication logs are missing, short-lived, or unqueryable. Most lateral movement still leaves footprints in identity systems: unusual logon types, impossible travel between VPN and cloud consoles, new MFA device…

Rehearsing Credential Rotation at Scale

Credential rotation at scale fails when it is treated as a heroic weekend project. Cloud keys, registry tokens, bot accounts, webhook secrets, and workload identities accumulate until a breach forces simultaneous rotation—and then production stalls because…

Hardening Internal Dev Platforms Against SSRF and Metadata Abuse

Internal developer platforms—artifact registries, webhook testers, preview environments, documentation importers, and MLOps-adjacent tools—often fetch URLs on behalf of users. That design convenience is also a classic server-side request forgery (SSRF) path to cloud metadata services and…

CI Control-Plane Patch Discipline for Build Servers

Build servers and CI control planes are crown jewels: they hold repository credentials, cloud deploy roles, signing keys, and the power to ship arbitrary artifacts into production. Critical remote-code-execution classes against CI products have repeatedly shown…

Privilege Access Workstations for Cloud Admins

Cloud administrator workstations are high-value targets. A single phishing email on a daily-driver laptop can yield console cookies, API keys, and enough session material to alter IAM or plant persistence. Privileged Access Workstations (PAWs)—hardened, dedicated environments…

ML Platform Security: Tracking Servers Should Not Be Public

Machine-learning platforms grew up as research convenience layers: tracking servers, model registries, feature stores, and notebook hosts that made experimentation easy. That convenience often arrived with default network exposure, weak authentication, and service identities that can…

Isolating Untrusted Code Execution in Internal Eval Platforms

Internal platforms increasingly execute untrusted or lightly trusted code: plugin sandboxes, customer-submitted jobs, research notebooks with shell access, and automated evaluation harnesses. Isolation is no longer optional. This evergreen guide covers practical boundaries for compute, identity,…