Security audit

huawei-cloud-ascendc-operator-performance-optim

Security checks across malware telemetry and agentic risk

Overview

This is a purpose-focused AscendC operator optimization skill, with local profiling scripts that need cautious use because they can remove prior profiling output.

Install only for AscendC/CANN operator development. Run the helper scripts from a copied or dedicated operator directory, preserve any OPPROF_* results you need before profiling, and avoid the chmod 777 /tmp/opprof troubleshooting command on shared machines; use a private per-user output directory instead.

SkillSpector

By NVIDIA

Vulnerability Patterns

Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
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 (6)

Context-Inappropriate Capability

Medium

Confidence: 93% confidence
Finding: The script unconditionally runs `rm -rf OPPROF_*` inside a user-supplied operator directory, which is a destructive action even if limited to profiling-output-like names. In this skill context, cleanup of old profiling artifacts is related to performance analysis, but deleting files without validation, scoping safeguards, or explicit consent can still remove unintended data if the directory contents are broader than expected.

Vague Triggers

Medium

Confidence: 86% confidence
Finding: The trigger list contains generic terms such as "performance", "custom operator", and "operator optimization" that can match many unrelated requests, causing the skill to activate outside its intended scope. Over-broad activation increases the chance the agent will use specialized tooling or domain guidance in inappropriate contexts, which can lead to incorrect actions or unsafe command suggestions.

Natural-Language Policy Violations

Medium

Confidence: 97% confidence
Finding: This document is a performance/operator API reference for low-level AscendC development, but the text is heavily machine-translated and ambiguous to the point that parameter meanings, limits, alignment requirements, and memory constraints can be misunderstood. In this context, unclear guidance can directly cause unsafe kernel implementations, including out-of-bounds accesses, corrupted results, or misuse of repeat/mask/count parameters on NPU hardware.

Missing User Warnings

Medium

Confidence: 97% confidence
Finding: The troubleshooting guidance recommends `chmod 777 /tmp/opprof`, which makes the profiling directory world-writable and world-accessible without any warning or compensating controls. On multi-user systems this can enable tampering, symlink attacks, or unauthorized access to profiling outputs, and `/tmp` is a particularly sensitive shared location for such misuse.

Missing User Warnings

Medium

Confidence: 90% confidence
Finding: The cleanup happens automatically with no confirmation, dry-run mode, or warning beyond a log message, so users may lose profiling data from previous runs unexpectedly. In an optimization/profiling skill this behavior is somewhat expected, which lowers suspicion of malice, but the lack of guardrails still makes it an unsafe destructive operation.

Missing User Warnings

Medium

Confidence: 95% confidence
Finding: The script unconditionally runs `rm -rf OPPROF_*` in the provided operator directory, which can delete existing profiling artifacts without confirmation, backup, or scoping to a newly created run directory. While the glob is narrower than a full path wipe, it still causes destructive data loss if valuable prior profiling outputs match that pattern.

VirusTotal

64/64 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.