INTERACTIVE INSIDER THREAT DETECTION DEMO
Run synthetic security scenarios through behavioral normalization, sequencing, and anomaly scoring.
A security engineering portfolio project by Vyshyvka Studio
Select an incident and commence the procedure
5-stage behavioral detection pipeline
Raw Vendor Logs
2021-11-08T09:12:04Z [celFS Audit Log]
{"action":"READ","path":"/models/alpha-a/params/decorrelation","user":"jwu","workstation":"QUANT-WS-114","value":"0.85","ticket":null}
OCSF Normalization
Behavioral Tokenization
Sequence Assembly
Narrative Analysis
09:12:04File:Read:Success:Info:AllowedResearcher reads decorrelation parameter for Model A — routine access
09:14:37File:Modify:Success:Info:AllowedParameter write to Model A decorrelation — value reduced significantly
09:31:02Process:Launch:Success:Info:System:AllowedModel A deployment via PAM — follows parameter modification
10:42:18File:Modify:Success:Info:AllowedSame parameter write pattern now on Model B — sequence repeating
11:05:44Process:Launch:Success:Info:System:AllowedModel B deploy — write→deploy pattern now seen twice for this identity
14:18:55File:Modify:Success:Info:AllowedThird model targeted — same celFS write pattern, decorrelation near-zero
14:44:31Process:Launch:Success:Info:System:AllowedThird write→deploy in sequence — behavioral frequency diverging from baseline
09:33:12File:Modify:Success:Low:AllowedFourth model — identical pattern is now statistically anomalous
09:55:48Process:Launch:Success:Low:System:AllowedFourth deploy — write→deploy→write→deploy is now a locked behavioral loop
11:07:22File:Modify:Success:Medium:AllowedBatch write across three models at once — acceleration of established anomalous pattern
11:38:09Process:Launch:Success:Medium:System:AllowedBatch deploy of three models — same identity, same pattern, scaling up
15:22:41Network:Traffic:Success:High:Unknown:AllowedModels showing 0.97 correlation where 0.15-0.35 expected — the mirror effect
What Happened
Quantitative researcher Jian Wu at Two Sigma systematically reduced decorrelation parameters to near-zero across 14 algorithmic investment models over 21 months. Models stopped generating independent alpha and began mirroring profitable positions, inflating his compensation from $2.75M to $23.5M. No system detected it. An anonymous social media post did.