Back to skill

Security audit

nim-interactive-chat

Security checks for vulnerabilities and agentic risk

Overview

The skill appears intended for local NIM chat setup, but it relies on Docker execution with an unpinned third-party image and generated scripts that may delete containers and leave services running.

Review the generated scripts before running them. Use a pinned and trusted container image, confirm what Docker privileges, mounts, ports, and environment variables are used, back up or rename any existing container with the same name, and avoid entering secrets into the chat unless you understand where the API traffic and logs go.

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:35
Finding

Unpinned Third-Party Container Image Creates a Supply-Chain Risk

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 35-42
Vulnerability Type: Unpinned third-party executable dependency
Risk Level: Medium

Vulnerable Code Snippet:

bash
create-nim-interactive \
  tgcr.turing-agi.com/public/qwen/qwen3.5-4b-spark-nim:latest \
  qwen3.5-4b-nim \
  18000 \
  Qwen/Qwen3.5-4B \
  0.88

Technical Analysis

The documented example instructs users to execute a container image obtained from tgcr.turing-agi.com using the mutable latest tag. A mutable tag does not provide an immutable association with the image that was originally reviewed or tested. The registry owner—or an attacker who compromises the registry, repository, publisher account, or release process—can replace the image referenced by this tag.

The project does not pin the image to a content digest, document image-signature verification, or provide provenance information. Consequently, the effective executable content can change without any corresponding modification to this repository.

The repository contains only SKILL.md; the referenced create-nim-interactive implementation and generated Docker scripts are absent. The audit therefore cannot determine which volumes, devices, capabilities, environment variables, or network permissions would be granted to the container.

Attack Path

  1. An attacker compromises the external registry, image repository, publisher credentials, or image build pipeline.
  2. The attacker publishes a malicious image under the existing qwen3.5-4b-spark-nim:latest tag.
  3. A user follows the documented workflow and invokes create-nim-interactive.
  4. The generated startup workflow pulls or runs the image currently associated with the mutable tag.
  5. Attacker-controlled code executes inside the container with all resources and permissions granted by the unreviewed startup script.
  6. Depending on those grants, the malicious image may access mounted files, GPU de ...[truncated 870 chars]
Remediation
View remediation

Remediation Suggestions

  1. Replace the mutable tag with an immutable digest, for example:
    bash
    trusted-registry.example/model/image@sha256:EXPECTED_DIGEST
    
  2. Prefer an authenticated, verified official NVIDIA NGC source or another explicitly trusted registry.
  3. Verify image signatures and provenance before execution, using mechanisms such as Cosign, Notary, or registry-native trust policies.
  4. Publish the expected digest and signature identity in the documentation and enforce verification in the generated startup script.
  5. Run the container with least privilege: avoid --privileged, the Docker socket, host namespaces, unnecessary capabilities, and broad writable host mounts.
  6. Use a non-root container user, a read-only root filesystem, capability dropping, restricted networking, and explicit resource/device grants where compatible.
  7. Include the create-nim-interactive implementation and generated-script templates in the project so command construction, shell quoting, container deletion behavior, mounts, environment variables, and Docker privileges can be audited.
  8. Add automated image vulnerability scanning and a controlled process for reviewing and updating pinned digests.
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
Findings (3)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The documentation explicitly advertises automatic deletion of old containers to avoid conflicts, but it does not warn users that an existing container with the same name may be removed and any non-persistent state lost. This is a real safety issue because users may run the generated script expecting a non-destructive startup flow and accidentally destroy running workloads or debugging state.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

The natural-language instructions throughout the file are presented only in Chinese, which can amount to a language policy issue when no user opt-in or locale justification is provided. The document does not indicate that the skill is intended exclusively for a Chinese-speaking or region-specific audience.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
80% confidence
Finding

The skill describes generating an OpenAI-compatible interactive chat client and connecting to a model service, but it does not disclose network/privacy implications such as prompts being sent over a local or remote API endpoint and possible logging of conversation data. While not inherently malicious, omission of this warning can mislead users about where sensitive inputs go and how they may be exposed.

Content

No source excerpt is available for this finding.