model-supply-chain-v1
22 rules · Keryn Software Licence · rules 0.8.0 · all packs
How to read a rule
- Severity
- How bad the finding is if it is real:
errorfails a default gate,warningandinfodo not. - Confidence
- P0 exploitable as written; P1 dangerous in context; P2 a heuristic or a posture check: confirm by reading the code.
- False-positive risk
- low: an exact API or flag, so hits are almost always real; medium: the pattern is legitimate in many programs, so check each hit; high: it finds a lead to confirm, so expect benign hits.
- Edge
- The relationship the finding sits on. reaches: content or traffic gets to a thing; calls: one thing invokes another; reads: one thing reads from another; writes: one thing writes to another; sends: data leaves for another party; grants: authority is given or left unchecked; depends_on: something relies on a component it does not verify; trusts: one party believes another without evidence. The boundary it crosses, when there is one: user->prompt (untrusted input reaching a prompt); model->code (model output reaching code, a shell, a query or a page); secret->outside (a secret leaving the trusted side); supply-chain (code, models or prompts pulled in from outside).
- Taxonomy chips
- References for readers who work in OWASP LLM 2025, OWASP Agentic, CWE or NIST AI RMF. They say where a rule belongs, not that a framework is satisfied.
seceng.modelsc.hf-revision-mutable-ref
Section titled “seceng.modelsc.hf-revision-mutable-ref”warning python huggingface, transformers
Hugging Face download pins revision= to a branch or tag name (e.g. “main”), which the repo owner can move at any time. Pin to the 40-character commit SHA instead.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
Fix: Resolve the reviewed commit once (huggingface_hub.HfApi().model_info(repo).sha) and pass that 40-hex SHA as revision=. Record it in config, bump it deliberately, and prefer safetensors weights with trust_remote_code left off.
References: https://huggingface.co/docs/huggingface_hub/guides/download · https://huggingface.co/docs/transformers/main_classes/model
seceng.modelsc.hf-from-pretrained-missing-revision
Section titled “seceng.modelsc.hf-from-pretrained-missing-revision”warning python huggingface, transformers
from_pretrained loads a Hub repo id with no revision=, so the weights resolve to whatever the default branch currently holds. Pin a reviewed 40-hex commit SHA.
Confidence P2: a heuristic or a posture check: confirm by reading the code. False-positive risk medium: the pattern is legitimate in many programs, so check each hit.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
Fix: Pass revision=<40-hex commit SHA> (and local_files_only=True in production once the snapshot is cached or baked into the image). Treat a repo-id literal without a pin as a floating dependency.
References: https://huggingface.co/docs/huggingface_hub/guides/download · https://huggingface.co/docs/transformers/main_classes/model
seceng.modelsc.hf-hub-download-missing-revision
Section titled “seceng.modelsc.hf-hub-download-missing-revision”warning python huggingface
hf_hub_download/snapshot_download fetches a repo with no revision=; the artifact floats with the default branch. Pin a 40-hex commit SHA.
Confidence P2: a heuristic or a posture check: confirm by reading the code. False-positive risk medium: the pattern is legitimate in many programs, so check each hit.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
Fix: Pass revision=<40-hex commit SHA>. For fully offline deploys download once in CI, verify, and bake the snapshot into the image or artifact store.
References: https://huggingface.co/docs/huggingface_hub/guides/download
seceng.modelsc.hf-inference-client-user-model
Section titled “seceng.modelsc.hf-inference-client-user-model”error python data-flow rule huggingface, flask, fastapi
Request-controlled value selects the model for huggingface_hub.InferenceClient. The model argument also accepts a URL, so callers can steer traffic (and your token) to an arbitrary endpoint or an unreviewed model. Map user choices to an allow-list of model ids.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: reaches content or traffic gets to a thing; it crosses the user->prompt boundary (untrusted input reaching a prompt).
CWE-918 CWE-829 LLM03 ASI04 ASI03
Fix: Never pass request data as model=. Keep a server-side dict of permitted model ids/endpoints, look the user’s choice up in it, and reject unknown keys. Do not attach a Hub token to requests whose destination the caller can choose.
References: https://huggingface.co/docs/huggingface_hub/guides/inference
Try it: a snippet that triggers this rule
Written for this page, not taken from any project. Save it as api/generate.py and scan it with keryn scan api --verbose; the finding below appears.
from fastapi import FastAPIfrom huggingface_hub import InferenceClient
app = FastAPI()
@app.post("/generate")def generate(model: str, prompt: str): client = InferenceClient(model=model) return {"text": client.text_generation(prompt, max_new_tokens=128)}Checked with Keryn’s browser engine when this page was built: it reports seceng.modelsc.hf-inference-client-user-model at line 9 (error, proof 2). The rule’s own logic is not published, so it cannot run in this tab; the pack list in the playground shows what does.
seceng.modelsc.transformers-user-controlled-model-id
Section titled “seceng.modelsc.transformers-user-controlled-model-id”error python data-flow rule huggingface, transformers, flask, fastapi
Request-controlled value is used as the Hugging Face model id (pipeline, from_pretrained, hf_hub_download, snapshot_download). An attacker can load any public repo, including one with pickled weights or custom code. Resolve ids through a server-side allow-list.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: reaches content or traffic gets to a thing; it crosses the user->prompt boundary (untrusted input reaching a prompt).
CWE-829 CWE-502 LLM03 ASI04 ASI05
Fix: Do not forward user input as a repo id or path. Look the choice up in a fixed mapping of reviewed models pinned by commit SHA, keep trust_remote_code off, and load safetensors only.
References: https://huggingface.co/docs/transformers/main_classes/pipelines · https://huggingface.co/docs/hub/security-pickle
seceng.modelsc.model-file-user-controlled-path
Section titled “seceng.modelsc.model-file-user-controlled-path”error python data-flow rule llama-cpp, gguf, safetensors, flask, fastapi
Request-controlled path reaches a local model loader (llama_cpp.Llama, LlamaCpp, safetensors load, GGUFReader). Callers can point the server at arbitrary files on disk or at attacker-supplied weights. Resolve a model name through an allow-list inside a fixed models directory.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: reaches content or traffic gets to a thing; it crosses the user->prompt boundary (untrusted input reaching a prompt).
Fix: Accept a model key, not a path. Map it to a file under a fixed directory (resolve and verify the result stays inside it), and verify a pinned SHA-256 of the file before loading.
References: https://llama-cpp-python.readthedocs.io/en/latest/api-reference/ · https://huggingface.co/docs/safetensors/index
seceng.modelsc.langsmith-pull-prompt-unpinned
Section titled “seceng.modelsc.langsmith-pull-prompt-unpinned”warning python langsmith, langchain
pull_prompt("name") fetches the latest version of a LangSmith prompt. Anyone with write access to the prompt can change what your application sends to the model; pin name:commit_hash.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
CWE-829 LLM03 LLM01 ASI04 ASI01
Fix: Pull by commit: client.pull_prompt(“owner/name:<commit_hash>”). Promote new commits through review, and store the hash in configuration.
References: https://docs.langchain.com/langsmith/manage-prompts-programmatically
seceng.modelsc.langchain-hub-pull-literal-unpinned
Section titled “seceng.modelsc.langchain-hub-pull-literal-unpinned”warning python langchain, langsmith
hub.pull("owner/prompt") has no :commit suffix, so the newest published prompt is used. A prompt edit upstream changes your agent’s instructions without a code change; pin the commit hash.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
CWE-829 LLM03 LLM01 ASI04 ASI01
Fix: Use hub.pull(“owner/prompt:<commit_hash>”) and keep the hash in version control; review prompt diffs before bumping.
References: https://docs.langchain.com/oss/python/langchain/overview · https://docs.langchain.com/langsmith/manage-prompts-programmatically
Try it: a snippet that triggers this rule
Written for this page, not taken from any project. Save it as agent/prompts.py and scan it with keryn scan agent --verbose; the finding below appears.
from langchain import hubfrom langchain_openai import ChatOpenAI
llm = ChatOpenAI(model="gpt-4o-mini")
def build_chain(): prompt = hub.pull("acme-team/support-triage") return prompt | llmChecked with Keryn’s browser engine when this page was built: it reports seceng.modelsc.langchain-hub-pull-literal-unpinned at line 8 (warning, proof 0). The rule’s own logic is not published, so it cannot run in this tab; the pack list in the playground shows what does.
seceng.modelsc.ollama-pull-latest-tag
Section titled “seceng.modelsc.ollama-pull-latest-tag”warning python ollama
Ollama model is pulled as latest (explicitly or by omitting the tag). The tag is mutable, so a rebuilt image or a repulled model can silently change the weights. Pin an explicit tag, ideally a digest recorded at review time.
Confidence P2: a heuristic or a posture check: confirm by reading the code. False-positive risk medium: the pattern is legitimate in many programs, so check each hit.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
Fix: Pull name:<specific-tag> and record the resulting digest (ollama show –modelfile / /api/tags) in your deployment manifest; avoid pulling at request time.
References: https://github.com/ollama/ollama/blob/main/docs/api.md#pull-a-model · https://github.com/ollama/ollama-python
seceng.modelsc.ollama-pull-user-model
Section titled “seceng.modelsc.ollama-pull-user-model”error python data-flow rule ollama, flask, fastapi
Request-controlled value names the model passed to Ollama’s pull. Callers can make the server download arbitrary (including oversized or attacker-hosted host/ns/model) artifacts. Allow-list model names.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: reaches content or traffic gets to a thing; it crosses the user->prompt boundary (untrusted input reaching a prompt).
CWE-829 CWE-770 LLM03 LLM10 ASI04
Fix: Pull only models from a server-side allow-list, pinned to a tag. Do not expose model management (pull/create/copy) to end users.
References: https://github.com/ollama/ollama/blob/main/docs/api.md#pull-a-model
Try it: a snippet that triggers this rule
Written for this page, not taken from any project. Save it as api/models.py and scan it with keryn scan api --verbose; the finding below appears.
import ollamafrom fastapi import FastAPI
app = FastAPI()
@app.post("/models/pull")def pull_model(name: str): ollama.pull(name) return {"status": "pulled", "model": name}Checked with Keryn’s browser engine when this page was built: it reports seceng.modelsc.ollama-pull-user-model at line 9 (error, proof 2). The rule’s own logic is not published, so it cannot run in this tab; the pack list in the playground shows what does.
seceng.modelsc.model-download-without-hash-check
Section titled “seceng.modelsc.model-download-without-hash-check”warning python requests, gguf, llama-cpp
Model weights (.gguf, .bin, .safetensors, .ckpt, …) are downloaded over HTTP in a function that never computes a digest. The file can be swapped upstream or in transit; verify a pinned SHA-256 before loading.
Confidence P2: a heuristic or a posture check: confirm by reading the code. False-positive risk medium: the pattern is legitimate in many programs, so check each hit.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
Fix: Hash the downloaded bytes with hashlib.sha256 and compare to a digest pinned in your repo (use hmac.compare_digest) before writing to the model directory or loading.
References: https://docs.python.org/3/library/hashlib.html · https://huggingface.co/docs/hub/security
seceng.modelsc.mcp-stdio-npx-unpinned-python
Section titled “seceng.modelsc.mcp-stdio-npx-unpinned-python”warning python mcp, npm
MCP stdio server is launched via npx -y <package> (or uvx) with no version, so every start executes whatever the registry currently serves as that package. Pin package@x.y.z (or an integrity-checked local install).
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
CWE-829 CWE-494 LLM03 ASI04 ASI02
Fix: Pin the exact version (npx -y @scope/server@1.2.3, uvx pkg==1.2.3) or install it at build time from a lockfile and launch the local binary.
References: https://modelcontextprotocol.io/specification/2026-07-28/basic/transports · https://docs.npmjs.com/cli/commands/npx
seceng.modelsc.langchainjs-hub-pull-unpinned
Section titled “seceng.modelsc.langchainjs-hub-pull-unpinned”warning typescript, javascript langchain, langsmith
hub.pull("owner/prompt") (langchain/hub) has no :commit suffix, so the newest published prompt is used. Pin the commit hash.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
CWE-829 LLM03 LLM01 ASI04 ASI01
Fix: Use hub.pull(“owner/prompt:<commit_hash>”) and keep the hash in version control.
References: https://docs.langchain.com/oss/javascript/langchain/overview
seceng.modelsc.ollama-pull-user-model-ts
Section titled “seceng.modelsc.ollama-pull-user-model-ts”error typescript, javascript data-flow rule ollama, express, nextjs
Request-controlled value names the model passed to ollama.pull. Callers can make the server download arbitrary artifacts; allow-list model names and pin tags.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: reaches content or traffic gets to a thing; it crosses the user->prompt boundary (untrusted input reaching a prompt).
CWE-829 CWE-770 LLM03 LLM10 ASI04
Fix: Pull only models from a server-side allow-list pinned to a tag; do not expose model management to end users.
References: https://github.com/ollama/ollama-js#pull
seceng.modelsc.transformersjs-user-controlled-model-id
Section titled “seceng.modelsc.transformersjs-user-controlled-model-id”error typescript, javascript data-flow rule transformers-js, huggingface, express, nextjs
Request-controlled value is the model id for Transformers.js pipeline/from_pretrained, letting callers load any Hub repo. Resolve ids through an allow-list.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: reaches content or traffic gets to a thing; it crosses the user->prompt boundary (untrusted input reaching a prompt).
Fix: Map user choices to a fixed list of reviewed model ids pinned with the revision option.
References: https://huggingface.co/docs/transformers.js/api/pipelines
Try it: a snippet that triggers this rule
Written for this page, not taken from any project. Save it as server/embed.js and scan it with keryn scan server --verbose; the finding below appears.
import express from "express";import { pipeline } from "@xenova/transformers";
const app = express();app.use(express.json());
app.post("/embed", async (req, res) => { const modelId = req.body.model; const extractor = await pipeline("feature-extraction", modelId); const out = await extractor(req.body.text, { pooling: "mean" }); res.json({ dims: out.dims });});
app.listen(3000);Checked with Keryn’s browser engine when this page was built: it reports seceng.modelsc.transformersjs-user-controlled-model-id at line 9 (error, proof 2). The rule’s own logic is not published, so it cannot run in this tab; the pack list in the playground shows what does.
seceng.modelsc.mcp-stdio-npx-unpinned-ts
Section titled “seceng.modelsc.mcp-stdio-npx-unpinned-ts”warning typescript, javascript mcp, npm
MCP stdio server is launched via npx -y <package> with no version, so each start runs whatever the registry currently serves. Pin package@x.y.z.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
CWE-829 CWE-494 LLM03 ASI04 ASI02
Fix: Pin the exact version (npx -y @scope/server@1.2.3) or install from a lockfile at build time and run the local binary.
References: https://modelcontextprotocol.io/specification/2026-07-28/basic/transports · https://docs.npmjs.com/cli/commands/npx
seceng.modelsc.dockerfile-ai-server-base-image-floating
Section titled “seceng.modelsc.dockerfile-ai-server-base-image-floating”warning dockerfile docker, vllm, ollama
Dockerfile base image for an AI serving stack (vLLM, Ollama, TGI, Open WebUI, LiteLLM) uses :latest/:main/:nightly or no tag. The build is not reproducible and an upstream push changes what you ship. Pin a version tag or digest.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
Fix: Use FROM image:<version>@sha256:<digest>; update via Renovate/Dependabot with review.
References: https://docs.docker.com/build/building/best-practices/#pin-base-image-versions
seceng.modelsc.dockerfile-pip-unpinned-agent-framework
Section titled “seceng.modelsc.dockerfile-pip-unpinned-agent-framework”warning dockerfile docker, python, langchain, openai
pip install in a Dockerfile installs an LLM/agent framework with no version specifier. Each rebuild pulls whatever is newest (including a hijacked or breaking release). Pin == versions or install from a hashed lockfile.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
Fix: Install from a lockfile: pip install –require-hashes -r requirements.lock (pip-compile –generate-hashes / uv pip compile), or at minimum pin pkg==x.y.z.
References: https://pip.pypa.io/en/stable/topics/secure-installs/
seceng.modelsc.dockerfile-npx-npm-global-unpinned-mcp
Section titled “seceng.modelsc.dockerfile-npx-npm-global-unpinned-mcp”warning dockerfile docker, mcp, npm
Dockerfile runs an MCP server through npx -y or npm install -g without a version. The container fetches whatever the registry serves at build or start time. Pin package@x.y.z.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
CWE-829 CWE-494 LLM03 ASI04 ASI02
Fix: npm install -g @scope/server@1.2.3 at build time (ideally from a lockfile with integrity hashes), then run the installed binary rather than npx -y.
References: https://docs.npmjs.com/cli/commands/npm-install · https://modelcontextprotocol.io/specification/2026-07-28/basic/transports
seceng.modelsc.dockerfile-ollama-pull-latest
Section titled “seceng.modelsc.dockerfile-ollama-pull-latest”warning dockerfile docker, ollama
Dockerfile pulls an Ollama model with latest or no tag. The tag moves, so rebuilds ship different weights. Pin an explicit tag.
Confidence P2: a heuristic or a posture check: confirm by reading the code. False-positive risk medium: the pattern is legitimate in many programs, so check each hit.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
Fix: Use ollama pull name:<specific-tag> and record the digest in the image labels.
References: https://docs.ollama.com/faq
seceng.modelsc.compose-ai-image-latest-or-untagged
Section titled “seceng.modelsc.compose-ai-image-latest-or-untagged”warning yaml docker-compose, kubernetes, vllm, ollama, litellm
AI serving image (vLLM, Ollama, TGI, Open WebUI, LiteLLM) in compose/Kubernetes YAML uses :latest/:main/:nightly or no tag. Deployments are not reproducible and an upstream push changes the running code. Pin a version tag or digest.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
Fix: Reference image:<version>@sha256:<digest>. Update through a reviewed bump (Renovate/Dependabot).
References: https://docs.docker.com/reference/compose-file/services/#image · https://kubernetes.io/docs/concepts/containers/images/#image-names
Try it: a snippet that triggers this rule
Written for this page, not taken from any project. Save it as docker-compose.yml and scan it with keryn scan docker-compose.yml --verbose; the finding below appears.
services: llm: image: ollama/ollama:latest ports: - "11434:11434" volumes: - ollama-data:/root/.ollama app: build: . environment: OLLAMA_HOST: http://llm:11434 depends_on: - llm
volumes: ollama-data: {}Checked with Keryn’s browser engine when this page was built: it reports seceng.modelsc.compose-ai-image-latest-or-untagged at line 3 (warning, proof 0). The rule’s own logic is not published, so it cannot run in this tab; the pack list in the playground shows what does.
seceng.modelsc.package-json-script-npx-unpinned-mcp
Section titled “seceng.modelsc.package-json-script-npx-unpinned-mcp”warning json npm, mcp
package.json script runs an MCP server with npx -y/npm i -g and no version. Each run executes whatever the registry currently serves. Pin package@x.y.z or add it to devDependencies with a lockfile.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: depends_on something relies on a component it does not verify; it crosses the supply-chain boundary (code, models or prompts pulled in from outside).
CWE-829 CWE-494 LLM03 ASI04 ASI02
Fix: Add the server to dependencies with an exact version, commit the lockfile, and invoke the local binary (npx –no-install server).
References: https://docs.npmjs.com/cli/commands/npx · https://modelcontextprotocol.io/specification/2026-07-28/basic/transports
Try it: a snippet that triggers this rule
Written for this page, not taken from any project. Save it as package.json and scan it with keryn scan package.json --verbose; the finding below appears.
{ "name": "docs-assistant", "version": "0.4.1", "private": true, "scripts": { "build": "tsc -p .", "mcp:files": "npx -y @modelcontextprotocol/server-filesystem ./docs", "start": "node dist/index.js" }, "dependencies": { "zod": "3.23.8" }}Checked with Keryn’s browser engine when this page was built: it reports seceng.modelsc.package-json-script-npx-unpinned-mcp at line 7 (warning, proof 1). The rule’s own logic is not published, so it cannot run in this tab; the pack list in the playground shows what does.