Security engineer, Berlin
Detection engineering, threat hunting, DFIR, purple teaming. Day to day I'm investigating intrusions and hunting for things that haven't alerted yet. Detections come out of both.
I run a lab at home and publish what comes out of it. Two things at the moment.
Telemetry fidelity. Run a known technique against an instrumented host, then compare what Sysmon, Wazuh and LimaCharlie each recorded. Most of the gaps turn out to be config rather than the tool.
Agent behaviour. Everyone is chasing malicious prompts. A prompt can be rephrased forever. The action can't, so that's what the rules watch.
Rules and an engine for spotting malicious AI agent behaviour in execution traces.
detection:
ingest:
step_type: tool_result
source_trust: untrusted
privileged_action:
step_type: tool_call
tool_name|in: ${privileged_tools}
condition: |
ingest followed by privileged_action within same session_id,
where privileged_action.tool_name not in tools_called_before_first_ingest
Sigma rules, and the telemetry comparison above.