Back to skill

Security audit

Gpu Container Setup Flagos

Security checks across malware telemetry and agentic risk

Overview

The skill mostly matches its GPU container setup purpose, but it tells the agent to permanently rewrite its own trusted image-source guidance based on web-discovered container registries.

Install only if you are comfortable supervising Docker pulls and GPU/data mounts. Use explicit trusted image names or known vendor registries when possible, and do not allow the skill to edit its own reference files based on web search without reviewing and approving the change.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (11)

Tp4

High
Category
MCP Tool Poisoning
Confidence
77% confidence
Finding
The skill description understates operational behavior by omitting data-disk discovery and overstates other actions, which can mislead users and reviewers about what will actually happen. In security-sensitive automation, description-behavior mismatch reduces informed consent and can hide risky filesystem inspection or future expansion into broader host manipulation.

Description-Behavior Mismatch

Medium
Confidence
95% confidence
Finding
The skill instructs self-modification of its reference files based on runtime-discovered information, but this behavior is not disclosed in the manifest. Hidden self-modification is dangerous because it changes future behavior and trust boundaries without review, creating a persistence mechanism for bad data or manipulated sources.

Context-Inappropriate Capability

Medium
Confidence
96% confidence
Finding
Web-driven self-improvement lets untrusted internet content influence persistent local configuration by promoting discovered registries into primary sources. That creates a supply-chain risk: a malicious or spoofed search result could become a trusted default for future container pulls, leading to persistent compromise.

Context-Inappropriate Capability

Medium
Confidence
95% confidence
Finding
The document explicitly authorizes unrestricted web searching to discover registries beyond a fixed allowlist. That expands the skill's operational scope from GPU container setup into open-ended retrieval of executable container sources, increasing the risk of selecting malicious or typo-squatted registries and undermining predictable behavior.

Description-Behavior Mismatch

High
Confidence
99% confidence
Finding
The guide instructs the agent to modify its own reference file after finding a working image, creating persistent self-modification not disclosed in the skill metadata. This allows runtime-discovered external content to become trusted future behavior, which is a classic persistence and supply-chain risk.

Missing User Warnings

Medium
Confidence
94% confidence
Finding
The skill plans to modify repository files as 'self-improvement' without warning the user that local files will be changed. Silent repository mutation is dangerous because it can persist unintended or attacker-influenced configuration changes and violates user expectations about a setup task.

Missing User Warnings

Medium
Confidence
98% confidence
Finding
The file tells the agent to persistently update the skill after successful web discovery without any explicit user consent or review checkpoint. Silent behavior changes are dangerous because users invoke a GPU setup skill expecting deterministic setup actions, not modification of future agent behavior based on untrusted network results.

Missing User Warnings

Medium
Confidence
98% confidence
Finding
The self-improvement rule mandates persistent edits as part of normal execution flow, without warning about the behavioral and supply-chain risks of modifying the skill over time. This creates hidden statefulness and makes future runs depend on prior network-discovered content that may have been malicious or erroneous.

Self-Modification

High
Category
Rogue Agent
Content
If test fails, try next source. If all fail, ask user for image.

#### Step 4.5: Update Skill (self-improvement)

**IMPORTANT**: If image found via Web Search (Step 4.3) passes all tests, update `references/image-sources.md` to add the newly discovered vendor hub as a primary source. This makes future lookups faster.
Confidence
98% confidence
Finding
Update Skill

Self-Modification

High
Category
Rogue Agent
Content
3. Query available tags
4. Select and pull image

### IMPORTANT: Update Skill After Success

If web search finds a working image that passes all tests (pull + PyTorch import + GPU detection), **you MUST update this file** to add the newly discovered registry as a primary vendor hub.
Confidence
99% confidence
Finding
Update Skill

Self-Modification

High
Category
Rogue Agent
Content
│         ↓ fail → Step 3                                         │
├─────────────────────────────────────────────────────────────────┤
│ Step 3: Web Search for vendor registry                          │
│         ↓ success → use image → UPDATE SKILL with new hub       │
│         ↓ fail → Step 4                                         │
├─────────────────────────────────────────────────────────────────┤
│ Step 4: Check local images, test PyTorch                        │
Confidence
98% confidence
Finding
UPDATE SKILL

VirusTotal

VirusTotal findings are pending for this skill version.

View on VirusTotal

Static analysis

No suspicious patterns detected.