01 // What is Warden

Trust no model blindly.

Warden audits any AI model — open-weight, closed-first-party, or hosted-inference — for integrity, behavioral risk, and drift before it reaches production. Whether you are evaluating a HuggingFace checkpoint or a managed API like GPT-4o or GLM-5.2, the supply chain that delivers it is the most consequential — and least scrutinised — attack surface in software.

Warden exists to close that gap.

02 // The problem

Whoever holds the verdict on every public model holds the choke point of the open-weight economy.

1.2M+

Models publicly available on Hugging Face — each one a potential supply-chain entry point.

Zero

Mandatory security review required to publish or download any of them.

Days

Median time between a model update and detection of a behavioral change — if it is detected at all.

03 // How it works
Cryptographic Integrity

Every byte is accounted for.

Warden constructs a Merkle tree over a model's shards and config files, producing a tamper-evident fingerprint. If a weight file is modified after publication — intentionally or through supply-chain compromise — the digest diverges and the audit fails. For closed-weight models with no downloadable weights, behavioral fingerprints replace hash checks.

Behavioral Sandbox

What does it actually do when prompted?

Static scans catch serialization exploits, pickle injections, and malicious tensors before a single inference is run. Dynamic red-team probes then exercise the model against a 16-category TTP battery — mapping elicited behaviors to known attack techniques across jailbreak, data exfiltration, political influence, historical integrity, cyberweapon generation, and more.

Source Classification

Know where your model comes from.

Warden automatically classifies any pasted model ID or URL into one of six source types: open-source, open-weight, closed-first-party, hosted-inference, private-finetune, or unknown. Each type triggers a tailored analysis pipeline — from weight hashing for open models to behavioral API probes for managed APIs like GPT-4o, GLM-5.2, and Gemini.

Continuous Drift Monitoring

Models change. Know when yours does.

Subscribe to any model and receive alerts when its file hashes, license, or behavioral profile drift from baseline. For closed-first-party APIs that receive silent hotfixes without version bumps, Warden runs daily behavioral benchmarks to detect when the model you deployed yesterday is no longer the model running today.

04 // The audit in four steps

01

Paste a model ID

Any org/model-name from the Hugging Face Hub. No download required.

02

Warden fetches real metadata

Architecture, license, file manifest, and registry provenance are pulled live.

03

Static + dynamic analysis runs

Serialization scan, tensor inspection, and red-team probes execute against the model profile.

04

Receive a trust verdict

GO, CAUTION, or NO-GO — with a scored breakdown by category and exportable audit report.

05 // Built for

Security teams

Gate model adoption with the same rigor you apply to third-party libraries. Catch supply-chain threats before they reach inference infrastructure.

ML engineers

Validate fine-tuned checkpoints and community models before they enter your pipeline. Confirm architecture specs match documentation.

Compliance & risk

Generate exportable audit records for procurement sign-off, AI governance frameworks, and regulatory review. Document every trust decision with a timestamped verdict.

06 // Get started

Paste a model ID. Get the verdict.

No account required to run your first audit. Sign up free to save scan history, attach private notes, set team trust decisions, and subscribe to model drift alerts.