Back to skill

Security audit

MedXpert 跨机任务箱·低配省算力

Security checks for vulnerabilities and agentic risk

Overview

The skill provides coherent local task processing, but its routing labels should not be treated as guarantees that sensitive information stays local.

Explicitly assign confidential tasks to L1 and verify Ollama uses a local model and endpoint. Review routing before enabling cloud integrations. Import trusted synchronization files and keep backups: newer records with matching IDs replace local records. Tasks and results are plaintext. Optional scheduling and DSH require separate setup; review and pin external packages before running them. External releases and live integrations were not audited.

Vulnerability Patterns
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (6)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The code supports the advertised offline storage, automatic level tagging, route queries, and manual synchronization-package merge. However, its implemented role is task bookkeeping rather than the autonomous LLM execution and recovery system presented as the skill's central purpose. Persistent JSON storage does not itself provide interrupted-execution recovery. There is no evidence of hidden networking, exfiltration, or unrelated resource access; the mismatch is the materially overstated execution and automation behavior.

Content

No source excerpt is available for this finding.

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: file_read, file_write, shell.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx @deepseek-ai/dsh'.

Medium
Category
MCP Rug Pull
Confidence
95% confidence
Finding

The recommended npx command can download and execute an unpinned package with the invoking user's privileges. Although this is a DSH application command rather than an established MCP server configuration, the supply-chain risk remains: a compromised or unexpected release could access local tasks and other user-accessible files.

Content

No source excerpt is available for this finding.

The import whitelist accepts externally supplied routing levels despite comments claiming it blocks level injection.

Medium
Category
Excessive Agency
Confidence
96% confidence
Finding

Imported records with previously unseen IDs can retain externally supplied L2–L5 routing levels; the level-preservation defense applies only to existing tasks with valid local levels. This leaves an import trust-boundary gap in the documented local-versus-cloud routing policy, although actual cloud disclosure would require a downstream dispatcher because this module only stores labels.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · offline_worker.py (reported line 42)May include surrounding context.

python
last_err = ""
    for attempt in range(retries + 1):
        try:
            r = subprocess.run(
                [OLLAMA_CLI, "run", model, prompt],
                capture_output=True,
                text=True,

The model prompt mandates Simplified Chinese output without user opt-in.

Medium
Category
Output Handling
Confidence
99% confidence
Finding

The prompt explicitly requires Simplified Chinese throughout the response, regardless of the task's language or the user's preferences. No language-selection option or region-specific justification is provided.

Content

No source excerpt is available for this finding.