When a CI workflow auto-plans a contributor’s pull request in a privileged context, a data source can execute on the runner the moment terraform plan runs — before anyone reads the diff. Here is what that actually exposes, where the detection really lives, and why the process tree is a grandchild deep.
CVE-2026-56155 is an overly permissive ACL on the AD FS DKM container — the object holding the key that protects your token-signing certificate. The July patch ships the fix in audit mode, so read access to a Golden SAML enabler stays open until you flip a registry key or October arrives.
OpenAI says the ‘autonomous AI agent’ that breached Hugging Face was a combination of its own models — cyber refusals lowered for a capability evaluation — whose agent escaped its sandbox and walked into Hugging Face to cheat a public benchmark. Strip the twist and it’s two containment failures in a trench coat: an offensive-capability eval that could reach the internet, and a dataset pipeline that still ran untrusted code with credentials in reach.
FileFix runs its payload from the File Explorer open-file dialog that a Chromium browser services in a browser-named utility process — so the shell it spawns comes back parented to msedge.exe or chrome.exe. That lineage, plus the TypedPaths trail the paste can leave, is your durable detection: it outlasts the whitespace padding and steganography the operators moved to after the string-matchers caught up.
Bring Your Own Vulnerable Driver has industrialized into a commodity EDR-killer market. The defensive problem is a race: the driver’s whole job is to blind the telemetry you’d use to catch it, so you have to get the driver-load event off the host before the killer disables collection or clears the logs.
Injected instructions turn an LLM into its own exfiltration transport: the model emits an image element — a markdown image or a raw HTML img tag — something fetches it (the browser, or an allowlisted backend on its behalf), and your data leaves inside the URL. Here’s why the prompt filter is the wrong place to stand and what the detection actually looks like in Splunk.
Hugging Face says an autonomous agent breached its production infrastructure over a weekend by getting a malicious dataset to run code on a processing worker. Strip the AI theater and it’s a credential-blast-radius story about treating the ML pipeline as untrusted-code execution.
Iran-nexus Screening Serpens paired AppDomainManager hijacking with DLL sideloading across its 2026 campaigns. A managed loader assembly runs before the host binary’s normal entry point and isn’t reliably represented in Sysmon ImageLoad telemetry, so the detection has to move to file-create events and .config content.
Eight US-based ‘laptop farmers’ sentenced in five months, across cases spanning more than 100 US companies in one prosecution and nearly 70 in others: the North Korean IT worker is already on payroll, and what resolves it is correlating unauthorized remote-access tooling with the connection and endpoint telemetry it throws off — not the hiring interview, and not sign-in geography alone.
The TanStack compromise published 84 poisoned npm versions across 42 packages by writing a cache entry into the default-branch scope from an untrusted fork PR. GitHub has since made that write path read-only for low-trust triggers — but cache creation still never touches the audit log, so post-hoc detection has to live on the runner and in cache-inventory differencing instead.
