Back to skill

Security audit

AI持续学习系统

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed autonomous learning logger, but it persists untrusted internet content into agent learning files and advertises long-term memory use without enough scoping or controls.

Review before installing. This skill is not clearly malicious, but it is designed for unattended internet collection and persistent agent learning. Only use it if you are comfortable with GitHub/arXiv data being summarized and saved locally, and require HTTPS arXiv URLs, prompt-injection handling, explicit retention controls, and a human review step before anything is promoted into long-term memory.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • 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
Findings (2)

T02 · Agent Memory Poisoning

Warning
Location
idle_learning.py:107
Finding

Untrusted Remote Metadata Can Poison Persistent Learning State

Content
View full analysis
{stars_min}" url = f"https://api.github.com/search/repositories?q={urllib.parse.quote(query)}&sort=stars&order=desc&per_page=3" try: import urllib.request req = urllib.request.Request(url, headers={"Accept": "application/vnd.github.v3+json"}) with urllib.request.urlopen(req, timeout=10) as resp: data = json.loads(resp.read()) items = data.get('items', [])[:3] if items: return [(i['full_name'], i['stargazers_count'], i.get('description','')) for i in items] except: pass return [] ``` ```python def arxiv_fetch(rss_url): """抓取arXiv论文""" import feedparser try: feed = feedparser.parse(rss_url) papers = [] for entry in feed.entries[:3]: papers.append({ 'title': entry.title.replace('\n','').strip(), 'summary': entry.summary[:200].replace('\n','').strip(), 'url': entry.id, }) return papers except: return [] ``` ```python def llm_summarize(topic, content): """轻量摘要 — 仅在内存充足时调用,否则跳过""" if not content: return None # 检查可用内存(MB) try: with open('/proc/meminfo', 'r') as f: for line in f: if line.startswith('MemAvailable:'): mem_available = int(line.split()[1]) / 1024 break except: mem_available = 500 if mem_available < 500: ...[truncated 4877 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
idle_learning.py:23
Finding

arXiv Feed Retrieval Uses Plaintext HTTP

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • 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 (15)

Tp4

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

The description overstates autonomous behavior such as hourly execution, nightly distillation, and long-term memory integration that are not actually enforced or implemented in the described file. This is dangerous because operators may grant trust, persistence, or elevated access based on claimed behavior, while hidden or undocumented runtime behavior such as local gating and local writes reduces transparency and auditability.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
92% confidence
Finding

The skill advertises and documents capabilities that imply file reads, file writes, network access, and shell execution, but it does not declare any explicit tool scope or permissions. This weakens least-privilege controls and can cause the skill to be deployed with broader access than users or the host expect, increasing the blast radius if the skill is modified, misused, or compromised.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill describes automatic scheduled collection from external sources and automatic writes to local files and memory without warning users about privacy, persistence, bandwidth, or operational impact. In unattended environments, this can lead to unexpected data accumulation, overwrite risk, or background network activity that users did not knowingly authorize.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The module description is written as an unconditional Chinese-language behavior statement, and later the summarization prompt explicitly instructs the model in Chinese. This indicates a fixed language/locale choice without any user opt-in or justification, which violates the language/locale policy criteria.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
89% confidence
Finding

The skill automatically transmits search queries and fetched learning content to external services on a schedule, creating a data egress channel. In this context the transmitted strings are mostly hardcoded topics, which lowers severity, but the autonomous background behavior and additional later forwarding of fetched content to another external LLM increase privacy, compliance, and supply-chain risk.

Content

Scanner excerpt · idle_learning.py (reported line 70)May include surrounding context.

python
import urllib.parse
    for kw in keywords:
        query = f"{kw} stars:>{stars_min}"
        url = f"https://api.github.com/search/repositories?q={urllib.parse.quote(query)}&sort=stars&order=desc&per_page=3"
        try:
            import urllib.request
            req = urllib.request.Request(url, headers={"Accept": "application/vnd.github.v3+json"})

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 · idle_learning.py (reported line 119)May include surrounding context.

python
summary_content = content[:150].replace('\n', ' ')
    prompt = f"用20字以内总结:{topic}。内容:{summary_content}"
    try:
        result = subprocess.run(
            ['/root/.nvm/versions/node/v22.22.2/bin/mmx', 'text', 'chat',
             '--model', 'abab6.5-chat', '--message', prompt, '--output', 'json'],
            capture_output=True, timeout=20

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The manifest describes an autonomous learning system that studies GitHub trends and arXiv papers and writes distilled results to long-term memory. While summarization is in scope, spawning an external executable through subprocess introduces a broader execution capability than the stated purpose requires; the same task could be performed without granting shell-adjacent process execution.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The manifest presents the skill as an automated learning and memory-distillation system, but these logged findings emphasize code-reading/modifying/verifying agent architectures. That content suggests a substantive interest in self-modifying/code-modifying agent capabilities not disclosed in the manifest’s stated scope.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The manifest describes an hourly continuous-learning system focused on GitHub trends and arXiv papers to improve AI capabilities. This entry labels the topic as "AI consciousness" but the findings include unrelated papers such as neutron star equations of state, indicating the skill is collecting content outside its stated AI-learning scope rather than focused self-improvement inputs.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

Under the topic "world model," the stored findings include papers on neutron stars and holography, which are not plausibly part of an AI self-improvement pipeline as described in the manifest. This shows behavior/data output drifting from the claimed domain-specific learning objective into generic or irrelevant paper collection.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
85% confidence
Finding

This markdown file contains many user-facing descriptions in Chinese, such as summaries and guidance text, while other entries remain in English. Because the file does not state that Chinese is required or provide any language/locale opt-in, it can violate a policy requiring user choice rather than forcing a specific language.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
80% confidence
Finding

The manifest describes an hourly system that learns from GitHub trends and arXiv papers, but this file shows the skill producing many synthesized topic buckets such as 'AI consciousness', 'self-aware AI architecture', 'hallucination detection', and 'world model', often with generated summaries that go beyond merely recording source materials. This indicates the skill is functioning as a broader autonomous topic-curation/summarization system rather than the narrower source-specific continuous learning behavior described in the manifest.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The natural-language content of the skill is entirely in Chinese, including the description and usage guidance, with no indication that users may choose another language or that the skill is intentionally restricted to a Chinese-speaking context. Under SQP-3, a forced language/locale without user opt-in can be a policy violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

This JSON file consistently contains natural-language summaries in Chinese while the referenced findings remain in English, indicating a language choice is being imposed in the artifact itself. The policy requires either user opt-in for a specific language/locale or a clearly documented justification, neither of which appears in this file.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
95% confidence
Finding

Under the 'world model' section, the text says the content is not actually related to world models and instead concerns unrelated subjects like karyotyping microservices and optimization. This is an active contradiction between the section label/intended categorization and the documented content.

Content

No source excerpt is available for this finding.