Interactive Evidence Dashboards

ML Security
Interactive Evidence

Interactive evidence dashboards for 13 ML security repositories. Selected benchmark results, detection metrics, and gap analysis from committed artifacts and test runs.

Honesty note:  These are interactive evidence dashboards, not live security monitoring. Most data is static benchmark output embedded in HTML. The MCP gateway is the only one with real-time capability when a server is running. Weak results are shown as-is — limitations are explained, not hidden.
Featured systems with scoped evidence
🛡 MCP Gateway Monitor

MCP Security Gateway Monitor

Inline MCP/JSON-RPC security gateway with prompt-injection patterns, capability checks, exfiltration signals, audit logging, rate limiting, and SIEM-oriented telemetry.

0
Passing tests
0%
Statement coverage
55
Injection patterns
View dashboard
🔍 Supply Chain Scanner

HF Model Provenance Scanner

Non-executing model-artifact scanner for pickle risk, provenance gaps, impersonation, suspicious loaders, serialization formats, and supply-chain indicators.

0
Passing tests
0%
Statement coverage
0
Attack fixtures
View dashboard
🧪 Prompt Injection Detector

LLM Redteam Framework

Offline prompt-injection detector evaluated on two datasets. F1=0.9714 on grouped held-out templates; novel-phrasing OOD F1=0.7188. Gap is documented, not papered over.

0.97
F1 grouped
0.7188
F1 novel OOD
−0.23
F1 drop
View dashboard
All 13 repositories

MCP Security Gateway Monitor

5-layer MCP tool-call firewall. 37/37 blocked on the fixed bundled self-test catalog; this is a regression signal, not a real-world detection rate.

Operational system 37/37 self-test

HF Model Provenance Scanner

Supply chain attack scanner for HuggingFace. Internal fixture suite: 12/12 detected. Not a general detection rate.

Operational system 12/12 fixtures

LLM Redteam Framework

Prompt-injection detector. F1=0.9714 grouped held-out / 0.7188 novel-phrasing OOD. Gap documented.

Operational system F1 0.97 → 0.7188 OOD

Adversarial ML Lab

FGSM/PGD/C&W attacks + Randomized Smoothing defense. CIFAR-10 SmallCNN measured: clean 71.82%, PGD-20 0.00% at ε=8/255.

Research / evaluation lab 71.82% → 0.00% (PGD-20)

Model Privacy Attacks

Membership inference + model extraction. Adult/OpenML DirectMIA mean AUC 0.557 across 5 seeds.

Research library MI AUC 0.557

Dataset Poisoning Detector

Synthetic label-flip benchmark: spectral detector average F1 0.6142 versus 0.1531 for the feature-space ensemble across 5/10/20% contamination.

Production-oriented system Spectral F1 0.6142 avg

PulseNet RUL Forecasting

ICS/OT research prototype using NASA C-MAPSS. FD001 validation records F1=0.54; committed inference benchmark records p99=13.242 ms over 500 samples. No production deployment is claimed.

ICS/OT prototype F1=0.54

ATT&CK v19 Core

MITRE ATT&CK v19 data models. Current repository verification reports 156 passing tests across 108 test functions in 11 files.

Core library 156 passing tests

Attack Detection Engine

5-source telemetry pipeline mapped to ATT&CK. The repository documents 42 rules but currently labels that count unverified; CI configuration sets a 53% minimum coverage gate.

Core library 5 detectors

AWS Agent Identity Guard

Static IAM policy linter for AI agent workloads. Checks wildcards, PassRole, trust boundary issues.

Production-oriented system Static analysis

Unified ML Security Platform

Authenticated integration control plane. Latest verified CI: 62 unit tests, 35 integration tests, 51.76% statement coverage. Local Compose uses contract stubs; no operated production deployment is claimed.

Integration control plane 62 unit + 35 integration

ML Security Benchmark Suite

Benchmark runner with HMAC-SHA256 signed evidence. Smoke tests only — not full production benchmarks.

Smoke tests Signed evidence

ML Security Command Center

Selected-project status summary from static evidence snapshots. Visualization only — not the full repository inventory or live monitoring.

Simulated data Visualization only