Back to skill

Security audit

中式智慧记忆引擎

Security checks for vulnerabilities and agentic risk

Overview

This long-term memory skill is purpose-aligned but should be reviewed because it stores sensitive personal profiles locally in plaintext with limited user controls and relies on an unpinned parent setup script.

Install only if you are comfortable with a local long-term memory profile that can retain personal conversation content, emotions, commitments, relationship metadata, and preference inferences in readable files under your home directory. Review the parent `mflow-memory` setup script before running it, and consider whether you need deletion, retention, encryption, and opt-in controls first.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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
  • 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
Findings (2)

T09 · Insecure Skill Coding Practices

Warning
Location
core/memory_store.py:75
Finding

Sensitive user and behavioral profile data is stored in plaintext

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
SKILL.md:25
Finding

Unpinned external parent Skill setup script creates a supply-chain execution risk

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (35)

Tp4

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

This mismatch includes an undeclared local file persistence behavior, which is materially security-relevant because it adds data retention and storage exposure not reflected in permissions. In a memory skill processing sensitive personal inferences, hidden persistence increases privacy risk, unauthorized retention, and difficulty for users to give informed consent.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

This mismatch includes an undeclared local file persistence behavior, which is materially security-relevant because it adds data retention and storage exposure not reflected in permissions. In a memory skill processing sensitive personal inferences, hidden persistence increases privacy risk, unauthorized retention, and difficulty for users to give informed consent.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

This mismatch includes an undeclared local file persistence behavior, which is materially security-relevant because it adds data retention and storage exposure not reflected in permissions. In a memory skill processing sensitive personal inferences, hidden persistence increases privacy risk, unauthorized retention, and difficulty for users to give informed consent.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

This mismatch includes an undeclared local file persistence behavior, which is materially security-relevant because it adds data retention and storage exposure not reflected in permissions. In a memory skill processing sensitive personal inferences, hidden persistence increases privacy risk, unauthorized retention, and difficulty for users to give informed consent.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill explicitly proposes storing sensitive personal inferences such as emotions, values, concerns, and trust-related judgments without clear privacy notice, consent, or limits. These are highly sensitive derived data points, and long-term retention of inferred psychological or behavioral attributes can harm users if exposed, misused, or surfaced unexpectedly.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

Building a long-term trust profile by comparing what users say against what they do creates a behavioral surveillance and profiling system. This is especially dangerous because it stores inferred authenticity, consistency, and compliance traits that can be wrong, biased, or highly sensitive, yet may influence future agent behavior toward the user.

Content

No source excerpt is available for this finding.

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · core/timing_sensor.py (reported line 208)May include surrounding context.

python
}
        
        rule = rules.get(content_type, (False, "默认不记住"))
        return rule
    
    def should_remind(self, topic: str, 
                     last_reminded: Optional[datetime] = None) -> tuple[bool, str]:

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill is framed as embedding fixed 'Chinese' behavioral traits and social norms into memory and response logic without indicating user choice or consent. In a system that stores long-term personal profiles and influences recommendations or responses, this can lead to biased behavior, cultural stereotyping, and covert personalization that users may not expect or want.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The README explicitly states that user memory, promises, relationships, and inferred 'wisdom' data are persistently stored under the user's home directory, but provides no privacy notice, retention limits, consent flow, or deletion guidance. For a long-term memory system that captures preferences, emotions, and relationship information, this creates a real privacy and security risk because sensitive personal data may be retained indefinitely and exposed to other local processes, backups, or shared environments.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
83% confidence
Finding

The skill declares no explicit tool scope or permissions while the analyzed behavior indicates file read/write capability. That creates a transparency and least-privilege problem: users and hosts cannot accurately assess what the skill may access or persist, especially for a memory system handling personal data.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill frames behavior around a specific cultural model and even refers to 'Chinese genes,' which risks stereotyping users and applying behavioral assumptions without consent. In a memory and profiling system, forced cultural framing can skew inference, personalization, and trust judgments in discriminatory or manipulative ways.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The instructions direct persistence of private complaints, emotions, concerns, and consistency-related information in plain language, increasing the sensitivity and readability of stored records. Plain-language retention of such disclosures raises the likelihood of privacy harm, secondary misuse, and accidental exposure because the data is immediately interpretable and difficult to sanitize once stored.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

Proactive promise tracking and future reminders monitor a user's commitments over time and can surface them later in ways the user may not expect. Without clear warning and consent, this creates a privacy and autonomy risk, especially when the system tracks deadlines, obligations, or potentially sensitive personal commitments.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The storage rules instruct broad retention of decisions, reasons, promises, preferences, and lessons learned, which can accumulate a detailed long-term dossier of personal and project information. In the context of a memory skill, that breadth makes the issue more dangerous because the feature is designed to normalize ongoing collection and persistence rather than narrowly scoped recall.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The code persistently stores arbitrary user content, emotion, values, and concerns via MemoryStore.save without any visible consent, disclosure, retention control, or sensitivity filtering at this entry point. In a long-term memory system designed to profile users, this creates a real privacy risk because users may unknowingly provide sensitive personal data that is later retained and reused.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The get_context method explicitly states it is used for dialogue injection and returns stored user context directly from memory. Without field-level filtering, sensitivity classification, or purpose ограничения, downstream prompt construction can expose personal data back into model responses or to other components, increasing privacy leakage risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The code persists raw user memories plus inferred emotional state, values, and concerns to local JSONL files with no consent flow, disclosure mechanism, retention control, or protection at rest visible in this file. Because this skill is explicitly a long-term memory system and stores highly sensitive profile data, unauthorized local access or downstream reuse can expose intimate personal information.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The save flow stores raw user content together with inferred personal-profile attributes and then separately stores emotion records, creating persistent linkage between identity, content, and sensitive inferences. In a memory skill, this materially increases the blast radius of any later disclosure, prompt reuse, local compromise, or debugging/logging access because highly revealing derived attributes are retained long-term alongside original text.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

get_context aggregates recent memories, emotion trends, pending promises, values, and concerns for automatic injection into each conversation turn. That design broadens exposure of sensitive historical data to any downstream prompt consumer, increasing the chance of inadvertent disclosure, over-sharing, prompt leakage, or cross-context resurfacing of personal information.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
85% confidence
Finding

The module header and embedded descriptions are entirely in Chinese and present the skill as culturally specific, but the file does not indicate that this language/locale is optional or limited to a justified region-specific use case. Under the policy, forcing a specific language or locale without opt-in is a natural-language policy concern.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The module-level documentation describes automatic promise recognition/registration and reminder tracking as core capabilities. In the implementation, register only accepts manually supplied content, and reminder-related methods merely compute booleans or return strings; there is no code that detects promises from conversation input or actually sends reminders.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The stated skill purpose is a long-term memory engine with rich relational and reasoning features, including commitment tracking as one aspect. This file implements only standalone local CRUD/statistics for promises and a basic trust score, with no knowledge-graph integration, emotional understanding, relationship scoring, or broader memory behavior implied by the manifest.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The module persists user promise content, deadlines, context, and derived trust statistics to disk under the user's home directory without any notice, consent, retention policy, or access-control hardening. In a memory skill, these records can contain sensitive personal commitments and behavioral history, so silent persistence increases privacy risk and the blast radius of local compromise or unintended data sharing.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The module description states a culturally specific relationship model as a general rule, using Chinese-language framing and assumptions about how people should be treated by relationship tier. There is no indication that users can choose this locale/cultural mode or opt into it, which can violate language/locale policy expectations.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The module silently creates a persistent directory under the user's home folder and stores relationship profiles, interaction counts, timestamps, and topics without any consent, notice, retention controls, or access protections. Because this skill is explicitly a long-term memory and relationship-profiling system, the persistence is central to its behavior, which increases the privacy risk: sensitive behavioral data can accumulate over time and be exposed to other local users, backups, or later components that read the file.

Content

No source excerpt is available for this finding.