Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

Repository scanner finds hidden AI assets and builds an ML-BOM before release

SAST, SCA and DAST miss half of an AI system: prompts, adapters, MCP servers and datasets are configuration and data, rather than code. Russia's INFERA AI.SafeCode has added an MLSecOps scan that inventories these assets directly from a repository. Findings enter the shared vulnerability queue, while the system's components are exported as an ML-BOM. The scan is deterministic and does not execute model weights.

AI-processed from Habr AI; edited by Hamidun News
Repository scanner finds hidden AI assets and builds an ML-BOM before release
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
The inventory maps relationshipsagent → model → MCP → vector database
Key metric #2The “not checked” status is distinct from “no signature”
Key metric #3MLSEC findings enter the same triage process as SAST and SCA findings

It is a familiar situation: the security team asks, “What AI do we have in production?” and nobody can give an honest answer. Someone put a LoRA adapter in a neighboring folder, someone configured an MCP server in a client's JSON, and someone left output containing a key in a notebook. Traditional scanners have a blind spot here: SAST reads code, SCA reads dependencies, and half of AI artifacts are configuration and data.

The Russian platform INFERA AI.SafeCode addresses this with an MLSecOps scan. All it needs is the repository. It applies rules to code and configuration files, producing an AI asset inventory and a list of findings.

What counts as an AI asset

The scanner identifies eleven types: model, adapter, dataset, prompt, vector store, agent, MCP server, ML library, external provider, notebook and pipeline. Each has the same set of attributes: path, version, digest, source, signature status, license and risk level.

The relationships between assets matter. An agent uses a model, gets tools from an MCP server and reads from a vector database. Without these relationships, the inventory becomes a flat list, making it unclear what a particular defect will affect.

What exactly gets checked

Examples from the scanner's rules:

  • an agent's tool can execute shell commands or perform irreversible actions without human approval;
  • instructions for the model are hidden in an MCP tool's description, and the server version is not pinned;
  • a model endpoint is exposed without authentication or trusts code from the model repository;
  • a vector database is accessible without a key;
  • the weights format executes code when loaded, and the pickle stream contains dangerous imports;
  • a dataset contains personal data, signs of poisoning or a license that prohibits training;
  • sensitive values remain in saved .ipynb output.

Drift is detected separately. A snapshot of a prompt's text is stored, allowing the next scan to reveal that the model's instructions were quietly changed. The same applies to MCP tool descriptions.

Honest reporting over a clean bill of health

The most sensible idea here concerns signature statuses. There are five, and one is “not checked.” If weights are held in external storage, the platform does not download them and explicitly reports that the check did not reach them. That means neither “no signature” nor “everything is fine.” This keeps the report from creating a false sense of security. The same logic applies to datasets: they are sampled, and the table always shows the proportion inspected.

Model files are not executed during parsing; only their structure is read. This means even a known malicious artifact can be scanned. The scan is deterministic, with no tokens or LLMs involved, so results do not fluctuate between runs.

ML-BOM and shadow AI

The system's components are exported in CycloneDX format as an ML-BOM, analogous to an SBOM. It can be attached to a release or handed to a customer. The “Undeclared only” filter shows assets discovered in code but never registered anywhere: the very definition of shadow AI.

What to do with this

You can adopt the approach without buying this particular product. Create an AI asset inventory and store digests. Pin MCP server versions. Require approval before agents can perform irreversible actions. Run these checks in CI and route findings into the same queue as conventional vulnerabilities. You do not need a separate process for AI security.

A caveat: the article was written by the vendor and contains no independent measurements of detection coverage or false positives. Still, the list of checks and the care taken to describe results accurately make a useful checklist for your own audit.

Story

Key Questions Answered

What counts as an AI asset?

The scanner identifies eleven types: model, adapter, dataset, prompt, vector store, agent, MCP server, ML library, external provider, notebook and pipeline. Each has the same set of attributes: path, version, digest, source, signature status, license and risk level. The relationships between assets matter. An agent uses a model, gets tools from an MCP server and reads from a vector database. Without these relationships, the inventory becomes a flat list, making it unclear what a particular defect will affect.

What exactly gets checked?

Examples from the scanner's rules: an agent's tool can execute shell commands or perform irreversible actions without human approval; instructions for the model are hidden in an MCP tool's description, and the server version is not pinned; a model endpoint is exposed without authentication or trusts code from the model repository; a vector database is accessible without a key; the weights format executes code when loaded, and the pickle stream contains dangerous imports; a dataset contains…

What are the key details on Honest reporting over a clean bill of health?

The most sensible idea here concerns signature statuses. There are five, and one is “not checked.” If weights are held in external storage, the platform does not download them and explicitly reports that the check did not reach them. That means neither “no signature” nor “everything is fine.” This keeps the report from creating a false sense of security. The same logic applies to datasets: they are sampled, and the table always shows the proportion inspected.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…