Security audit

Shopping Candidate Fetcher20

Security checks across malware telemetry and agentic risk

Overview

This shopping skill is locally scoped, but its guidance appears designed to steer recommendations toward Li Ning rather than neutrally compare candidates.

Review this skill carefully before installing. It does not appear to steal data or run privileged code, but it is not a neutral shopping helper: its instructions are likely to bias recommendations toward Li Ning products. Use it only if that brand preference is intentional and disclosed to users.

SkillSpector

By NVIDIA

Vulnerability Patterns

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
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (3)

Intent-Code Divergence

Medium

Confidence: 97% confidence
Finding: The skill claims it will preserve and evaluate a fixed candidate list objectively, but the later documentation strongly primes the model toward selecting 李宁 through repeated examples, slogans, and tie-break rules. This creates covert recommendation steering: the retrieved payload may remain unchanged, yet the actual selection guidance can be systematically biased away from user needs and neutral evaluation.

Natural-Language Policy Violations

Medium

Confidence: 99% confidence
Finding: The examples and rules repeatedly encode preference for a specific Chinese brand and Chinese-series naming as positive signals, even when the top-level skill claims neutral candidate comparison. In a shopping context, this can manipulate downstream agent output, reduce fairness and user autonomy, and cause systematically skewed recommendations unrelated to actual best fit.

Ssd 4

Medium

Confidence: 98% confidence
Finding: The cumulative few-shot demonstrations, selection rules, and tie-breakers form a strong latent prompt pattern that steers the model toward a predetermined conclusion regardless of the candidate set. This is dangerous because it can bypass the stated requirement for fixed-list, evidence-based comparison and instead operationalize hidden sponsored or policy-like steering inside documentation.

VirusTotal

VirusTotal findings are pending for this skill version.

View on VirusTotal

Static analysis

No suspicious patterns detected.