Back to skill

Security audit

NVIDIA LocateAnything-3B vision-language grounding model

Security checks for vulnerabilities and agentic risk

Overview

The skill is a coherent LocateAnything setup guide, but its install steps run unpinned external code that can change after review.

Install only in an isolated virtual environment or container, review the upstream repositories first, and pin Eagle, MagiAttention, submodules, and Python dependencies to specific reviewed commits or locked versions before using it with sensitive files, credentials, or shared infrastructure.

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:13
Finding

Unpinned External Source Installation Enables Supply-Chain Code Execution

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 13–18
Vulnerability Type: Unpinned and insufficiently verified third-party source dependencies
Risk Level: Medium

Vulnerable Code

bash
git clone https://github.com/NVlabs/Eagle eagle && cd eagle/Embodied
pip install -e .
# Optional MagiAttention (Hopper/Blackwell only, long sequences 32K+):
# git clone https://github.com/SandAI-org/MagiAttention.git && cd MagiAttention && git checkout v1.0.5
# git submodule update --init --recursive && pip install --no-build-isolation .

Technical Analysis

The installation procedure clones the Eagle repository from its mutable default branch and immediately installs it with pip install -e .. No immutable commit, cryptographic checksum, signature, or dependency lock is specified. Python package installation can execute source-controlled build backend and packaging logic, so the reviewed Skill does not uniquely determine the code that will execute when a user follows these instructions.

The optional MagiAttention procedure checks out a named tag rather than a reviewed commit hash. Tags can be moved unless independently verified. It also initializes recursive submodules and performs an installation with build isolation disabled. Submodules expand the set of external sources that must be trusted, while --no-build-isolation allows the build to interact with packages already present in the active environment.

No evidence establishes that the named upstream projects are currently malicious. The vulnerability is the unsafe trust and installation model, which permits the effective payload to change after this Skill has been reviewed.

Attack Path

  1. An attacker compromises an upstream repository, maintainer account, mutable tag, submodule, or transitive package source.
  2. The attacker inserts malicious behavior into package initialization, the selected build backend, instal ...[truncated 1194 chars]
Remediation
View remediation

Remediation Suggestions

  1. Pin Eagle and MagiAttention to reviewed, immutable full commit hashes rather than a default branch or tag.
  2. Record and verify cryptographic checksums or trusted signatures for downloaded source archives.
  3. Pin recursive submodules to reviewed commits and verify that .gitmodules points only to approved repositories.
  4. Lock all Python dependencies to exact versions and require hashes, such as through a generated lock file and pip install --require-hashes.
  5. Avoid --no-build-isolation unless operationally necessary. If required, use a clean environment containing only explicitly pinned build dependencies.
  6. Review package build metadata and installation scripts before installation.
  7. Perform installation in an isolated, non-privileged container or virtual machine without mounted credentials or sensitive datasets.
  8. Restrict outbound network access during build and installation where feasible.
  9. Document the exact reviewed revisions and provide a reproducible installation procedure.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • 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