Pixeltable

• Outcome: Review
• Latest audit: May 28, 2026 · 10:29 AM UTC
• Version: 2.3.0

Overview

This is a coherent Pixeltable documentation skill, but it teaches high-impact AI, tool, memory, and deletion patterns without enough safety and data-handling guidance.

Review before installing. Use the recipes only with approved providers and trusted MCP servers, avoid private or regulated data unless you have reviewed vendor terms and retention, keep API keys in environment variables or a secret manager, require human approval for tools that can write, delete, spend money, or make network calls, and add auth, audit logs, backups, and retention/deletion policies before deploying the serving or memory examples.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access
• Supply Chain: Unpinned Dependencies, External Script Fetching, Obfuscated Code
• Excessive Agency: Unrestricted Tool Access, Autonomous Decision Making, Scope Creep

Findings (13)

Missing User Warnings

Medium

• Confidence: 88% confidence
• Finding: The skill prominently encourages use of external AI providers and tool-calling capabilities, but it does not clearly warn that prompts, documents, images, audio, and tool parameters may be transmitted to third-party services or trigger side effects. In a skill meant to guide agent behavior, this omission can lead users to unknowingly expose sensitive data or invoke actions through external systems.

Missing User Warnings

Medium

• Confidence: 94% confidence
• Finding: The agent pipeline example shows that inserting a row automatically triggers model calls, tool selection, tool execution, retrieval, and final response generation, but it does not include any guardrails around tool approval or action scope. This is dangerous because an agent or user following the example could wire in powerful tools and cause unintended external actions simply by inserting data.

Missing User Warnings

Medium

• Confidence: 86% confidence
• Finding: The provider integration section lists many third-party AI services and local/runtime integrations without any privacy, security, credential-handling, or data residency warning. Because the skill is instructional, the absence of such guidance increases the chance that users will route sensitive content to inappropriate providers or assume equivalent trust boundaries across services.

Missing User Warnings

Medium

• Confidence: 88% confidence
• Finding: The Tool Use section presents a pattern where an LLM can select tools and Pixeltable will execute them automatically, but it does not include guardrails, confirmation requirements, permission boundaries, or warnings about sensitive data and side effects. In an agent-building skill, this is dangerous because readers may generalize the pattern to real tools with network, filesystem, database, or transaction capabilities, creating unsafe autonomous action paths.

Missing User Warnings

Medium

• Confidence: 91% confidence
• Finding: The recipe persistently stores chat history and long-lived memories, which can include sensitive personal, operational, or proprietary data, but it does not include any retention, consent, minimization, or disclosure guidance. In a production agent-memory pattern, this omission can lead to unnecessary collection and prolonged storage of sensitive data, increasing privacy, compliance, and breach impact.

Missing User Warnings

Medium

• Confidence: 94% confidence
• Finding: The MCP guidance encourages connecting to arbitrary external MCP servers without warning that tool calls may receive user prompts, retrieved memories, system prompts, and other context. In this skill's context, MCP tools are integrated directly into an agent with persistent memory, so enabling a remote server can exfiltrate highly sensitive cross-session data to third parties if the server is untrusted or over-privileged.

Missing User Warnings

Medium

• Confidence: 74% confidence
• Finding: The serving section shows autogenerated delete endpoints that can remove records via an API call, but it does not mention authentication, authorization, confirmation, or audit controls. In a skill focused on production serving patterns, omission of these safeguards can normalize exposing destructive endpoints directly, increasing the chance of unsafe deployment.

Missing User Warnings

Medium

• Confidence: 90% confidence
• Finding: The examples send `images.image` to external model providers via `chat_completions()` using image URLs, but the documentation does not warn users that image contents may be transmitted off-platform to third-party AI services. In an ML data pipeline context, datasets often contain sensitive, proprietary, or regulated media, so users could unknowingly expose private data during labeling or captioning workflows.

Missing User Warnings

Medium

• Confidence: 96% confidence
• Finding: This provider reference prominently documents many integrations that transmit prompts, images, audio, video, embeddings, and other user-supplied content to external AI services, but it does not clearly warn users about data egress, retention, provider-side processing, or privacy/compliance implications. In a skill explicitly intended for production multimodal AI pipelines, omission of this warning can cause operators to unknowingly send sensitive data to third parties.

Missing User Warnings

Medium

• Confidence: 91% confidence
• Finding: The WhisperX diarization example notes that an HF token is required, but it does not explain that diarization can involve external model access and processing of sensitive biometric/voice data, including speaker segmentation metadata. Users may enable this feature on confidential audio without understanding the additional privacy and compliance risk.

Missing User Warnings

Medium

• Confidence: 97% confidence
• Finding: The pipeline sends video frames, audio, transcripts, prompts, and potentially derived context to external providers including OpenAI, Anthropic, Hugging Face-backed models, and DuckDuckGo search without an explicit warning in the recipe. In a multimodal RAG workflow, this can expose sensitive media content, transcribed speech, and user questions to third parties, creating privacy, compliance, and data-governance risk.

Missing User Warnings

Medium

• Confidence: 91% confidence
• Finding: The examples send user-provided or uploaded documents, images, audio, video frames, and prompts to external providers such as OpenAI, Anthropic, Together, DuckDuckGo, and others, but the documentation does not warn that this may transmit sensitive data off-system. In a production-oriented workflow guide, omission of privacy and data-handling warnings can lead developers to unintentionally exfiltrate regulated, proprietary, or personal data when they copy these patterns.

Missing User Warnings

Medium

• Confidence: 84% confidence
• Finding: The FastAPIRouter documentation shows `router.add_delete_route(docs, path="/delete")` as a generated endpoint but does not clearly warn that this exposes row deletion functionality over HTTP. Readers may enable a destructive route by copy-pasting the example without considering authentication, authorization, audit logging, CSRF protections, or accidental data loss.

VirusTotal

VirusTotal findings are pending for this skill version.

Static analysis

No suspicious patterns detected.