Back to skill

Security audit

Wsl2 Local Ai

Security checks for vulnerabilities and agentic risk

Overview

This skill is coherent for setting up local AI on WSL2, but its primary setup asks users to execute mutable remote installers and unpinned packages without verification.

Review the install commands before using this skill. Prefer pinned package versions, a tagged or digest-pinned Docker image, and downloading any installer for inspection and checksum or signature verification before execution. The local API examples are expected for this tool, but the setup changes your WSL2 environment and may start long-running local services.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • 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
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (3)

T03 · Remote Payload Retrieval and Execution

Error
Location
SKILL.md:41
Finding

Unverified Remote Installer Is Downloaded and Executed Directly

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 41
Vulnerability Type: Remote payload retrieval and immediate shell execution
Risk Level: High

Vulnerable Code:

bash
curl -fsSL https://ollama.ai/install.sh | sh

Technical Analysis

This command streams a mutable script from an external URL directly into sh. The effective executable payload is therefore determined at installation time rather than at Skill review time. No version pin, cryptographic signature, checksum verification, local inspection step, or content allowlist is present.

Installing Ollama is relevant to the Skill's declared functionality, but executing an unverified network response is not the minimum-risk installation method. HTTPS provides transport protection but does not protect against compromise of the vendor infrastructure, domain, publishing credentials, or the installer itself. It also does not guarantee that future content at the URL matches the version reviewed during this audit.

Depending on how the remote installer operates, it may modify system paths, install binaries or services, and request elevated privileges. Any malicious commands delivered by the endpoint would run with the privileges of the user invoking sh, including elevated privileges if the installer or user invokes sudo.

Attack Path

  1. An attacker compromises the installer endpoint, vendor publishing process, DNS/TLS trust path, or another component capable of controlling the returned script.
  2. The user follows the Skill instructions and runs the documented command.
  3. curl retrieves the attacker-controlled response.
  4. The pipe sends the response directly to sh without verification or review.
  5. The attacker-controlled commands execute with the invoking user's privileges.
  6. If elevated authorization is available or requested, the payload may install system-wide files, services, or other persistent components.

Impact Assessm

...[truncated 466 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace direct curl | sh execution with official manual installation instructions or a pinned release artifact.
  2. Download the artifact to a local file without executing it:
    bash
    curl --proto '=https' --tlsv1.2 -fL -o ollama-installer.sh \
      https://ollama.ai/install.sh
    
  3. Verify a vendor-published cryptographic signature or checksum obtained through an independently authenticated channel.
  4. Pin the expected release version and integrity value rather than trusting mutable latest content.
  5. Allow the user to inspect the downloaded script before explicitly executing it.
  6. Run installation with ordinary user privileges wherever possible. Clearly identify any step requiring elevation and explain why it is necessary.
  7. Prefer a trusted package manager or signed vendor package when one is available.

T08 · Insecure Dependencies

Warning
Location
SKILL.md:49
Finding

Unpinned Python Package Installation in Primary Setup Workflow

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 49
Vulnerability Type: Unpinned third-party dependency installation
Risk Level: Medium

Vulnerable Code:

bash
pip install ollama-herd

Technical Analysis

The installation command does not constrain ollama-herd to a reviewed version and does not verify package hashes. The package index can consequently resolve a release published after this Skill was audited. Python source distributions and build backends may execute code during package construction or installation, while installed console entry points execute package code when subsequently invoked.

The package is directly related to the declared fleet-routing functionality, so the dependency itself is functionally justified. However, selecting mutable latest content without integrity verification exceeds the minimum supply-chain trust necessary to install a known, reviewed version.

Attack Path

  1. The package publisher account, package index, release process, or a dependency is compromised, or a malicious future release is published.
  2. The user runs the unpinned pip install ollama-herd command.
  3. pip resolves and downloads the currently selected release rather than a previously audited version.
  4. Malicious build hooks may execute during installation, or malicious package code executes when the documented herd or herd-node commands are launched.
  5. The payload gains access to files, credentials, environment variables, and network services available to the invoking user.

Impact Assessment

Exploitation can result in arbitrary code execution with the WSL2 user's privileges. Potentially exposed assets include source repositories, Python credentials, SSH material, environment secrets, local model services, fleet configuration, and Windows files mounted into WSL2. System-wide impact is possible if the command is run from an elevated shell or against a privileged Python insta ...[truncated 8 chars]

Remediation
View remediation

Remediation Suggestions

  1. Pin ollama-herd to a specific version that has been reviewed:
    bash
    python3 -m pip install 'ollama-herd==<reviewed-version>'
    
  2. Publish a lock or requirements file containing hashes and install it with:
    bash
    python3 -m pip install --require-hashes -r requirements.txt
    
  3. Pin and hash transitive dependencies where practical.
  4. Use a dedicated virtual environment rather than a system Python installation.
  5. Prefer binary wheels from trusted sources and review any source distribution or custom build backend before installation.
  6. Document the expected package index and avoid untrusted extra indexes.

T08 · Insecure Dependencies

Warning
Location
SKILL.md:120
Finding

Unpinned Python Package Installation Repeated in Docker Workflow

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 120
Vulnerability Type: Unpinned third-party dependency installation
Risk Level: Medium

Vulnerable Code:

bash
pip install ollama-herd

Technical Analysis

The Docker-oriented workflow repeats the unrestricted installation of the latest package version without a lock file, version constraint, signature, or package hash. This creates the same mutable supply-chain execution boundary as the primary installation workflow. The surrounding use of Docker does not isolate this command because it is documented as running directly in WSL2 rather than inside the shown Ollama container.

A malicious package release or compromised build dependency can therefore execute in the host WSL2 user context during installation or when the installed herd commands are started.

Attack Path

  1. An attacker gains control over a package release, publisher account, package distribution channel, or dependency.
  2. The user follows the Docker workflow and executes the unpinned package installation on the WSL2 host.
  3. pip retrieves the attacker-controlled version selected at that time.
  4. Malicious installation hooks or later package entry points execute.
  5. The payload accesses the user's WSL2 environment and may interact with local Docker and Ollama services.

Impact Assessment

The payload obtains the permissions of the invoking WSL2 user. In addition to user files and credentials, access to a Docker socket or equivalent Docker control interface could allow control over containers and may provide a path to broader host access. The exact scope depends on the user's Docker permissions and whether installation is performed with elevated privileges.

Remediation
View remediation

Remediation Suggestions

  1. Remove the duplicate installation instruction and reference a single hardened installation procedure.
  2. Pin the package and all required dependencies to reviewed versions.
  3. Require package hashes through a locked requirements file and --require-hashes.
  4. Install into a dedicated virtual environment with ordinary user privileges.
  5. Do not grant the routing package Docker control unless that access is explicitly required.
  6. Document package provenance, supported versions, and a controlled update process.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Rogue AgentSelf-Modification, Session Persistence
  • YARA SignaturesMalware Match, Webshell Match, Cryptominer Match
Findings (9)

External Script Fetching

High
Category
Supply Chain
Confidence
98% confidence
Finding

Piping a remotely fetched script directly into sh executes unverified code from the network without inspection, integrity verification, or version pinning. If the upstream host, CDN, DNS, or connection is compromised, the user could execute arbitrary code inside WSL2 immediately.

Content

Scanner excerpt · SKILL.md (reported line 41)May include surrounding context.

bash
# Inside WSL2
curl -fsSL https://ollama.ai/install.sh | sh
ollama serve &

Chaining Abuse

High
Category
Tool Misuse
Confidence
99% confidence
Finding

The | sh chaining pattern is dangerous because it turns network content into immediately executed shell commands with no review boundary. In a setup guide, this materially increases the chance of supply-chain compromise affecting all users who follow the instructions.

Content

Scanner excerpt · SKILL.md (reported line 41)May include surrounding context.

bash
# Inside WSL2
curl -fsSL https://ollama.ai/install.sh | sh
ollama serve &

YARA rule 'backdoor_persistence': Backdoor persistence with malicious payloads (shell commands, SSH key injection, hidden root users) [malware]

High
Category
YARA Match
Confidence
75% confidence
Finding

YARA rule matched a known malware signature (reverse shell, backdoor, ransomware, C2 framework, or info stealer).

Content

Scanner excerpt · SKILL.md (reported line 141)May include surrounding context.

desktop | 16GB shared with WSL2 | phi4, codestral, qwen3.5:14b | | RTX 4060 laptop | 8GB shared with WSL2 | phi4-mini, gemma3:4b |

WSL2 shares GPU memory with Windows. Close GPU-heavy Windows apps for more WSL2 AI vRAM.

WSL2 AI environment

bash
# WSL2 Ollama optimization
export OLLAMA_KEEP_ALIVE=-1
export OLLAMA_MAX_LOADED_MODELS=-1

# Add to ~/.bashrc for persistence in WSL2
echo 'export OLLAMA_KEEP_ALIVE=-1' >> ~/.bashrc
echo 'export OLLAMA_MAX_LOADED_MODELS=-1' >> ~/.bashrc

Monitor WSL2 AI

bash
# WSL2 fleet status
curl -s http://localhost:11435/fleet/status | python3 -m json.tool

# WSL2 health checks
curl -s http://localhost:11435/dashboard/api/health | python3 -m json.tool

Dashboard at http://localhost:11435/dashboard — accessible from both Windows browser and WSL2.

Also available on WSL2 AI

Image generation

bash
curl http://localhost:11435/api/generate-image \
  -d '{"model": "z-image-turbo", "prompt": "developer workspace", "

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · SKILL.md (reported line 149)May include surrounding context.

bash
# WSL2 fleet status
curl -s http://localhost:11435/fleet/status | python3 -m json.tool

# WSL2 health checks
curl -s http://localhost:11435/dashboard/api/health | python3 -m json.tool

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · SKILL.md (reported line 152)May include surrounding context.

curl -s http://localhost:11435/fleet/status | python3 -m json.tool

WSL2 health checks

curl -s http://localhost:11435/dashboard/api/health | python3 -m json.tool

text

Dashboard at `http://localhost:11435/dashboard` — accessible from both Windows browser and WSL2.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 60)May include surrounding context.

powershell
# From Windows PowerShell
curl http://localhost:11435/api/tags    # see WSL2 AI models

Use WSL2 AI

Rp1

Medium
Category
MCP Rug Pull
Confidence
75% confidence
Finding

Docker image references without a specific tag (:latest is implicit) or digest (@sha256:...) can be silently replaced by a malicious image.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
90% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · SKILL.md (reported line 140)May include surrounding context.

export OLLAMA_KEEP_ALIVE=-1 export OLLAMA_MAX_LOADED_MODELS=-1

Add to ~/.bashrc for persistence in WSL2

echo 'export OLLAMA_KEEP_ALIVE=-1' >> ~/.bashrc echo 'export OLLAMA_MAX_LOADED_MODELS=-1' >> ~/.bashrc

text

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 161)May include surrounding context.

Image generation

bash
curl http://localhost:11435/api/generate-image \
  -d '{"model": "z-image-turbo", "prompt": "developer workspace", "width": 1024, "height": 1024}'