Back to skill

Security audit

Exo Cluster

Security checks for vulnerabilities and agentic risk

Overview

This skill is coherent for setting up an Exo local AI cluster, but it asks users to run high-impact installation and network service commands without enough security scoping.

Review this skill before installing. Use it only if you are comfortable running third-party installers and starting LAN-accessible inference services. Prefer pinned releases or reviewed commits, use a dedicated virtual environment or unprivileged account, restrict dashboard/API access to trusted hosts, and avoid sending sensitive prompts unless you have authentication, network isolation, and transport protections in place.

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 (2)

T03 · Remote Payload Retrieval and Execution

Error
Location
SKILL.md:50
Finding

Unverified Remote Shell Script Downloaded and Executed

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 50
Vulnerability Type: Remote payload retrieval and execution
Risk Level: Critical

Vulnerable Code:

bash
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

Technical Analysis

The command retrieves the current HEAD version of a remotely hosted shell script and immediately passes its contents to Bash. The remote payload is mutable because it is referenced through a branch rather than an immutable commit. The Skill performs no checksum validation, signature verification, version pinning, or local review before execution.

Although installing Homebrew is relevant to preparing the declared Exo environment, executing an unverified remote installer is not the minimum-privilege or minimum-risk installation method. The downloaded script runs with all permissions available to the invoking user and may request elevated privileges during installation.

Attack Path

  1. An attacker compromises the upstream repository, a maintainer account, or the referenced branch.
  2. The attacker modifies install.sh to include malicious shell commands.
  3. A user or agent follows the Skill and executes the documented command.
  4. curl downloads the modified payload without integrity validation.
  5. Bash immediately executes the attacker-controlled commands.
  6. The payload can access user data, alter the environment, install additional software, or attempt privilege escalation.

Impact Assessment

Successful exploitation provides arbitrary code execution under the invoking user's account. The payload can read or modify files available to that account, access environment variables and locally available credentials, install additional components, and alter development tooling. If the installer obtains administrative authorization, the impact may extend to system-wide files, packages, and configuration.

Remediation
View remediation

Remediation Suggestions

  • Do not pipe network content directly into a shell.
  • Prefer directing users to Homebrew's official installation documentation or using an already installed package manager.
  • If automated installation is required, pin the installer to a reviewed immutable commit.
  • Download the installer as a separate file and verify its cryptographic checksum or trusted signature before execution.
  • Review the downloaded script locally before running it.
  • Execute installation with an unprivileged account and grant elevated privileges only for individually reviewed operations.
  • Document the expected filesystem and permission changes so users can make an informed security decision.

T08 · Insecure Dependencies

Error
Location
SKILL.md:56
Finding

Unpinned Repositories and Dependencies Permit Supply-Chain Code Execution

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 56–57, 63–70, 77–78, and 98–105
Vulnerability Type: Insecure dependencies and mutable third-party source
Risk Level: High

Vulnerable Code:

bash
git clone https://github.com/exo-explore/exo
cd exo
bash
git clone https://github.com/tensorsofthewall/exo_windows
exo_windows
bash
git clone https://github.com/exo-explore/exo
cd exo
pip install -e .
python -m exo.main --role master
bash
git clone https://github.com/exo-explore/exo
cd exo
bash
nix run .#exo
bash
pip install -e .
pip install torch tinygrad

Technical Analysis

The repository clones use mutable default branches rather than reviewed release tags or commit hashes. Consequently, the code installed and executed may change after this Skill has been audited. The community Windows fork expands the supply-chain trust boundary beyond the declared upstream Exo project without pinning or integrity verification.

pip install -e . installs code from the newly cloned repository and may execute package build hooks or backend logic. The torch and tinygrad dependencies are also installed without version constraints or hashes. nix run .#exo evaluates and executes the Nix definition in the current mutable checkout.

These operations are relevant to installing Exo, but their mutable and unverified form exceeds the minimum supply-chain risk necessary for the declared functionality.

Attack Path

  1. An attacker compromises an upstream repository, maintainer account, community fork, or dependency release process.
  2. The attacker inserts malicious code into the default branch, package build configuration, Nix definition, or a newly resolved package version.
  3. A user follows the Skill and clones the current repository or resolves the latest package versions.
  4. pip install -e ., nix run .#exo, or direct Python ex ...[truncated 902 chars]
Remediation
View remediation

Remediation Suggestions

  • Pin each Git repository to a reviewed commit hash or cryptographically signed release tag.
  • Verify release signatures or published checksums before installation.
  • Clearly identify the Windows repository as a third-party community fork and document its separate trust boundary.
  • Pin Python package versions and use a lock file containing cryptographic hashes.
  • Install Python dependencies inside a dedicated virtual environment rather than the user's global environment.
  • Review package build metadata and installation hooks before running pip install -e ..
  • Prefer non-editable installation from a verified release artifact for routine users.
  • Review and pin the Nix flake inputs and lock file before running nix run.
  • Run the software as a dedicated unprivileged account and restrict access to credentials, personal files, and sensitive network resources.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (5)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The trigger phrases are very broad and could cause the skill to activate in loosely related conversations about local model deployment or clustering. That increases the chance of unsolicited operational guidance being injected into a session, including network/service setup steps that expose local APIs or cluster services without the user's explicit intent.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
94% confidence
Finding

The Homebrew installation command fetches a remote script with curl and executes it directly in the shell. This is dangerous because it grants full code execution to content retrieved over the network; if the source is compromised, intercepted, or replaced, the user may run arbitrary attacker-controlled code.

Content

Scanner excerpt · SKILL.md (reported line 50)May include surrounding context.

md
xcode-select --install

# 2. 安装 Homebrew
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

# 3. 安装依赖
brew install uv node rust

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill instructs users to start a master node, workers, a dashboard, and API-compatible endpoints accessible over the network, but it does not warn about authentication, binding scope, firewall rules, or sensitive prompt/data handling. In context, this can lead users to expose inference APIs and cluster control surfaces on a LAN or beyond, creating unauthorized access and data leakage risks.

Content

No source excerpt is available for this finding.

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 248)May include surrounding context.

示例:

bash
# 让集群运行推理
curl -X POST http://mac-server:52415/v1/chat/completions \
  -d '{"model": "deepseek-7b", "messages": [{"role": "user", "content": "解释量子计算"}]}'

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
80% confidence
Finding

The natural-language instructions, headings, trigger phrases, and examples are all presented in Chinese, effectively imposing a single language on users. The file does not offer an alternative language, user opt-in, or a documented justification for the locale constraint.

Content

No source excerpt is available for this finding.