topic
Prompt Injection
4 posts tagged “Prompt Injection”.
Stop reading me the attacker's words: fencing log-substrate prompt injection inside the Sentinel data path
LogFence is a Microsoft Sentinel DCR transform that contains log-substrate prompt injection at ingestion — fencing attacker-controlled fields before any AI reads them. The missing Microsoft implementation of an agreed idea, measured deterministically and on live telemetry.
What your DCR transform actually does to untrusted data (and why it matters now)
A CCF-authoring deep-dive: ingestion-time transform design quietly decides what your detections — and now your AI readers — actually see. What a transform really does to attacker-controlled text, the restricted-KQL gotchas that bite, and a non-destructive trust-tagging pattern you can adopt today.
Once a prompt injection reaches the AI, does it obey? I tested that too
Part 2 of a defensive-research series: once a planted instruction reaches an AI SOC-triage assistant, how often does it obey? Measured across two models and four defence conditions — with the finding that instructing the reader beats fancy ingestion tags, and that tags can backfire on cheap models.
Do prompt injections survive the Microsoft Sentinel pipeline? I measured it.
A defensive-research experiment: if an attacker hides an instruction inside a log field, does it survive a real Microsoft Sentinel ingestion pipeline all the way to where an AI assistant would read it? The answer — and why the connector's transform is the real control surface.