Back to skill

Security audit

Ollama — Herd Your LLMs Into One Smart Endpoint

Security checks for vulnerabilities and agentic risk

Overview

This skill is a coherent Ollama fleet-routing helper, with expected but important privacy and supply-chain considerations.

Install only in an isolated Python environment or with a pinned reviewed package version. Treat prompts, files, audio, images, metadata, and outputs as visible to any trusted fleet node that may handle the request, and confirm before allowing model downloads because they can consume substantial bandwidth and disk space.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (1)

T08 · Insecure Dependencies

Warning
Location
SKILL.md:16
Finding
Unpinned Third-Party Package Installation## Vulnerability Details **File Location**: `SKILL.md`, line 16 **Vulnerability Type**: Unpinned third-party dependency installation **Risk Level**: Medium **Vulnerable Code Snippet**: ```bash pip install ollama-herd ``` ### Technical Analysis The setup instructions install `ollama-herd` directly from the default Python Package Index without specifying an exact version or validating package hashes. Because the executable package source is not included in the audited project, its installation behavior and runtime implementation cannot be verified from this artifact. An unpinned installation resolves to whichever release the package index serves at installation time. If the publisher account, package distribution process, package index, or a future release is compromised, users following these instructions could install attacker-controlled code. Python package installation may execute package-controlled build logic, and the installed commands (`herd` and `herd-node`) subsequently execute with the privileges of the invoking user. This finding does not establish that the current `ollama-herd` package is malicious. It identifies a supply-chain weakness caused by mutable dependency resolution and the absence of integrity verification. ### Attack Path 1. An attacker compromises the package publisher account, release pipeline, or another component of the package distribution channel. 2. The attacker publishes a malicious release under the expected `ollama-herd` package name. 3. A user or agent follows the setup instruction and runs `pip install ollama-herd`. 4. Pip resolves and downloads the attacker-controlled release because no version or hash is constrained. 5. Malicious build or installation logic executes, or malicious behavior runs when the user starts `herd` or `herd-node`. 6. The payload operates with the privileges and accessible resources of the user who performed the installation or launched the installed commands. ...[truncated 541 chars]
Remediation
## Remediation Suggestions - Pin the dependency to a reviewed, exact version rather than installing the latest mutable release. - Publish a locked requirements file containing approved cryptographic hashes, and install it with pip's `--require-hashes` option. - Example hardened workflow: ```bash python3 -m venv .venv . .venv/bin/activate python3 -m pip install --require-hashes -r requirements.txt ``` - Review the selected source release and its build configuration before approving its hashes. - Run the router and node agents under a dedicated, non-privileged account with access limited to required files and network destinations. - Avoid installing the package with root privileges or into the system Python environment. - Use an internally controlled package mirror or vendored, reviewed artifact where stronger supply-chain assurance is required. - Add an upgrade process that requires explicit review, integrity verification, and testing before changing the pinned version.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (4)

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The skill markets a single local endpoint but does not clearly warn users that prompts and inputs may be routed to multiple machines and that operational data may be written to local fleet-manager files. In a multi-node LLM setup, this changes the trust boundary and can expose sensitive prompts, metadata, or outputs to additional hosts and logs without informed user consent.

External Transmission

Medium
Category
Data Exfiltration
Content
```bash
# Chat
curl http://localhost:11435/api/chat -d '{
  "model": "qwen3:235b",
  "messages": [{"role": "user", "content": "Hello"}],
  "stream": false
Confidence
60% confidence
Finding
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The documented auto-pull behavior means the system may download models automatically, causing unannounced network activity and local storage changes. Without a clear warning and opt-in, users may trigger large downloads, consume bandwidth, alter disk state, and fetch artifacts from external sources unexpectedly.

External Transmission

Medium
Category
Data Exfiltration
Content
### Image generation

```bash
curl -o image.png http://localhost:11435/api/generate-image \
  -H "Content-Type: application/json" \
  -d '{"model":"z-image-turbo","prompt":"a sunset","width":1024,"height":1024,"steps":4}'
```
Confidence
60% confidence
Finding
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Static analysis

No suspicious patterns detected.