ml_security_agent / v0.1.0

Your ML pipeline
has a blind spot.

Prism watches your ML pipeline 24/7 — catching data poisoning, model drift attacks, and training data leakage before they ship. Built for the engineers who've seen it happen.

prism monitor --live

2026-06-09 14:23:01 OK Pipeline healthy — no anomalies detected

2026-06-09 14:23:04 WARN Data drift detected: feature_dist_x (+4.2σ) in training batch

2026-06-09 14:23:07 ALERT Model version f7a2: inference anomaly — possible injection attempt

2026-06-09 14:23:10 BLOCKED Deployment halted. Artifact quarantined.

prism> _

Data Ingestion

Training

Validation

Deploy

Prism

watching

what_we_catch

Attacks that slip past every other tool

Data Poisoning

Malicious training data that subtly degrades your model — invisible to standard monitoring, catastrophic in production.

injection vectors

Model Drift Attack

Adversarial inputs that look normal but slowly shift your model's behavior — until your fraud detector misses real fraud.

concept shift

Training Data Leakage

PII or sensitive data making it into your training set — leaking through to predictions, violating compliance.

gdpr / soc2

Artifact Tampering

A compromised dependency swaps your model mid-deploy. Prism cryptographically verifies every artifact.

supply chain

"Security teams don't watch ML pipelines. ML teams don't know security. The gap between those two is where your model breaks, your data leaks, and your users pay the price."

We built Prism because we lived this problem. At Netflix, we ran ML infrastructure at scale — and watched security tooling miss every attack that mattered to our models. CVE databases don't flag a poisoned training batch. SIEMs don't catch when your feature distribution drifts toward an adversarial input.

Prism is built by ML infrastructure engineers who also understand attack surfaces. It speaks the language of your pipeline — and watches for the threats that only a practitioner would recognize.

We catch what the CVE databases miss.

No dashboards for the security team — real-time alerts for the engineers shipping models.

Your pipeline, not ours. Prism integrates into your existing stack — Metaflow, Airflow, Vertex, SageMaker.

Stop flying blind on your most critical infrastructure.

Every week your ML pipeline runs without monitoring is a week an attack could be quietly degrading your models. Prism starts watching in under five minutes. No infrastructure changes required.