Single-turn prompt injection usually dies when the context window clears. Memory poisoning persists, fires weeks later, and hides in a memory-write event your platform was never built to emit as security telemetry. Here is what to instrument, what to detect, and what the first round of tuning gets wrong.
An unauthenticated file-write in Splunk Enterprise’s bundled Postgres sidecar chains to code execution as the splunk service account — the same account that owns your _audit and _internal indexes. The detection problem is mostly a log-forwarding hygiene problem you either solved last year or didn’t.
Attackers don’t need a vulnerable driver to blind an EDR — they need the agent’s own installer and a window. The durable detection isn’t the kill command, it’s the silence that follows. Here is what that detection looks like the first time you deploy it, and why it floods the SOC before it works.
Keyless OIDC federation killed long-lived AWS keys in CI, but it moved the trust boundary into a JSON condition block most teams wrote once and never reread. Here is where the sub-claim hole lives, what CloudTrail can and can’t tell you, and why this is a configuration audit before it is ever a detection.
An attacker with app-management rights doesn’t need another password. They add their own secret to a trusted service principal and authenticate as the app, outside MFA and outside Conditional Access. Here is what the audit event looks like, what the first round of tuning has to fix, and where the cleanest preventive control hides behind a SKU you probably didn’t buy.
On ESXi the logs that explain how an intruder got root are written to a ramdisk that a reboot erases. Here is where the evidence actually lives, what to detect before it’s gone, and how it maps to the audit-storage controls the architecture violates by default.
Device code phishing produces a clean, MFA-satisfied sign-in on Microsoft’s own infrastructure — and most of the telemetry that betrays it sits in the Entra non-interactive log table teams drop to save money. Here’s where the detection actually lives, how the threshold flips between a Windows shop and a dev-heavy tenant, and the persistence artifacts the closeout always skips.
Encryption-type detancy deton Kerberos TGS requests still catches lazy roasting, but as RC4 gets disabled the durable signal moves to behavior. What to grep for, why the AES downgrade tell degrades, and how to kill service-account false positives before they bury the SOC.
A clean netstat is not proof of safety: eBPF backdoors like LinkPro keep an internal listener, rewrite ports through XDP/TC, and can make bpftool lie about themselves. The durable signal is the bpf() syscall at load time, and on Cilium-heavy fleets, telling real loaders from noise is most of the work.
Token Protection and DBSC closed some doors in 2026 but left the SOC with a messier detection surface than the vendor decks suggest. Here is what the incident actually looks like in the SIEM, where the first round of tuning has to land, and what most teams get wrong on revocation.
