Back to skill

Security audit

智能记忆触发系统

Security checks for vulnerabilities and agentic risk

Overview

This skill is documentation-only and not malware, but it encourages broad task-history review, interaction recording, durable workflow storage, and cross-agent sharing without clear consent or access boundaries.

Install only if you are comfortable with an agent being encouraged to inspect task history, record workflow-related interactions, save generated workflow documents, and potentially share derived workflows across agents. Before use, require explicit opt-in for history review, saving, and sharing; define storage location, retention, redaction, and per-user or per-workspace access controls.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (1)

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
usage-guide.md:120
Finding
Uncontrolled Collection and Cross-Agent Sharing of User Interaction Data## Vulnerability Details **File Locations**: - `SKILL.md:126` - `SKILL.md:175` - `memory-trigger-management.md:272-275` - `trigger-logic.md:11` - `usage-guide.md:69-73` - `usage-guide.md:120-124` **Vulnerability Type**: Excessive access to historical user data and insufficiently controlled cross-agent sharing **Risk Level**: Medium ### Vulnerable Snippets `SKILL.md:126`: ```markdown - **Weekly**: Check task history, identify high-complexity tasks without workflows ``` `SKILL.md:175`: ```text Check Task History (is it a repetitive type?) ``` `memory-trigger-management.md:272-275`: ```markdown ### With Agent Systems - **Task Execution**: Integrate with agent task execution systems - **Knowledge Sharing**: Share workflows across agent teams - **Collaboration**: Enable collaborative workflow creation ``` `trigger-logic.md:11`: ```text Check Task History (is it a repetitive type?) ``` `usage-guide.md:69-73`: ```markdown ### Medium-term Improvements (1 month) 1. **Machine Learning Integration**: Use ML to improve pattern recognition 2. **Cross-Agent Learning**: Share learning across different agents 3. **Automated Workflow Generation**: Automatically generate workflow drafts ``` `usage-guide.md:120-124`: ```markdown ### Data Collection Phase - Collect task execution data - Record user interactions - Track workflow usage - Gather user feedback ``` ### Technical Analysis The Skill instructs an agent to inspect task history, record user interactions and task-execution data, create persistent workflow documentation, and share workflows or learned information across agents. These operations require access beyond the immediate task context and may cross user, session, or agent isolation boundaries. The documentation does not define: - Explicit user consent before historical data is inspected or recorded. - A least-privilege scope for history access. - Separati ...[truncated 2873 chars]
Remediation
## Remediation Suggestions 1. Require explicit, informed opt-in before reading task history, recording interactions, or enabling cross-agent sharing. 2. Limit history access to the current user, workspace, and minimum time range required for repetition detection. 3. Use non-sensitive metadata or irreversible task fingerprints for similarity detection instead of retaining complete conversation content. 4. Add automatic redaction for credentials, tokens, personal data, internal addresses, configuration secrets, and other sensitive values before workflow creation. 5. Apply per-user, per-workspace, and per-agent access controls to all workflow documents and monitoring records. 6. Disable cross-agent sharing by default. Require a separate authorization decision for every destination agent or explicitly approved trust group. 7. Define retention periods, deletion procedures, and user controls for inspecting, correcting, exporting, and deleting recorded information. 8. Encrypt persisted workflow and monitoring data at rest and protect it in transit where sharing is explicitly authorized. 9. Record auditable events for history access, workflow creation, sensitive-data redaction, and cross-agent sharing. 10. Document that workflow generation must exclude raw secrets even when users explicitly request task documentation. 11. Add negative tests confirming that one agent, user, or workspace cannot retrieve workflows or interaction records belonging to another authorization scope.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • 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
Findings (17)

Vague Triggers

Medium
Confidence
94% confidence
Finding
The mandatory trigger conditions include broad, common signals such as tasks with 3+ steps and generic words like 'remember', 'summarize', and 'process'. In an agent skill, this can cause unintended activation on ordinary requests, leading the agent to proactively create or suggest documentation when the user did not clearly consent. Because the skill is explicitly designed to shift behavior from passive to active, the overbroad triggers materially increase the chance of scope creep and unwanted data capture.

Vague Triggers

Medium
Confidence
90% confidence
Finding
The recommended trigger conditions use vague concepts such as 'potential reuse', 'team collaboration', and 'error risk' without operational boundaries. This ambiguity gives the agent wide discretion to activate proactive behavior inconsistently, which can result in over-collection of task details or intrusive prompts in contexts the user did not intend. In a memory/documentation skill, unclear activation criteria are dangerous because they expand when the system may retain or formalize user activity.

Vague Triggers

Medium
Confidence
96% confidence
Finding
The task identification keywords are highly generic terms like 'create', 'edit', 'delete', 'run', and 'integration', which appear in many benign requests. This makes the trigger scope excessively broad and prone to false activations, potentially causing the agent to inspect history, classify tasks as repetitive, or propose recording workflows for routine activity. Given the skill's automation goal, generic keyword matching increases the risk of unintended monitoring and persistence behavior.

Description-Behavior Mismatch

Medium
Confidence
91% confidence
Finding
The manifest focuses on automatically determining when to create workflow documentation. These sections define maintenance, updates, version control, archiving, and retirement processes, which are broader document-management operations than the claimed trigger-oriented purpose.

Description-Behavior Mismatch

Medium
Confidence
93% confidence
Finding
The manifest describes an intelligent trigger system whose purpose is to determine when workflow documentation should be created and to proactively help. This file substantially expands that scope into performance management, experimentation, predictive workflow creation, analytics dashboards, reporting, and broad integrations, which are separate operational capabilities rather than a direct implementation detail of trigger determination.

Context-Inappropriate Capability

Medium
Confidence
88% confidence
Finding
Running control/experimental groups to optimize system behavior is an experimentation capability, not an obvious requirement for deciding when to create workflow documentation. The manifest does not declare experimentation or model-tuning infrastructure as part of the skill's purpose.

Context-Inappropriate Capability

Medium
Confidence
85% confidence
Finding
Integration with documentation platforms, version control systems, and monitoring tools introduces broader platform connectivity than is necessary to determine when workflow documentation should be created. The manifest does not state that the skill should connect to external systems or manage cross-platform integrations.

Natural-Language Policy Violations

Medium
Confidence
93% confidence
Finding
The note "Upgraded to English Version" indicates a language choice being enforced at the system level, but the document does not offer user opt-in, alternatives, or a documented locale-specific justification. This can violate language/locale policy requirements for skills that should not force a specific language by default.

Vague Triggers

Medium
Confidence
95% confidence
Finding
The activation criteria are broad and heuristic-driven, using generic signals like complexity, repetition, and intent keywords to decide whether to trigger memory behavior. In practice, this makes accidental invocation likely, which is dangerous in a skill that may ask to document or store operational details derived from ordinary conversation.

Vague Triggers

Medium
Confidence
96% confidence
Finding
The implicit intent detector treats common phrases such as 'how to' and 'can you show me' as signals for workflow or memory creation intent, even though they are ordinary requests for assistance. This overbroad matching can cause the skill to infer retention/documentation intent where none exists, leading to unnecessary prompting or downstream persistence of user task content.

Intent-Code Divergence

Medium
Confidence
94% confidence
Finding
The decision matrix conflicts with the stated hard requirements, creating ambiguous behavior around when workflows should be created versus when the user should merely be asked. In a memory-trigger skill, this inconsistency can cause the agent to persist information or initiate documentation in situations the policy did not clearly authorize, increasing the risk of unintended data capture and user-surprising behavior.

Intent-Code Divergence

Medium
Confidence
91% confidence
Finding
The storage logic assumes documents will be written directly to disk, which conflicts with the documented ask-before-create behavior and bypasses a clear consent checkpoint. Because the skill handles workflow memories, this can lead to silent persistence of task details onto the user's filesystem without an explicit, contemporaneous approval step.

Missing User Warnings

Medium
Confidence
96% confidence
Finding
The skill specifies concrete filesystem paths for saving generated workflow documents but does not require a user-facing warning or consent flow about persistence. This is especially risky because workflow documents may contain configuration details, procedural steps, or other sensitive operational context that would be silently written to a durable local location.

Missing User Warnings

Medium
Confidence
92% confidence
Finding
The document explicitly instructs agents to collect task execution data, user interactions, workflow usage, and user feedback, but it provides no privacy boundaries, consent requirements, retention limits, or guidance on handling sensitive content. In a memory/monitoring skill, this omission can normalize over-collection of potentially personal or confidential data and lead to unauthorized storage or secondary use.

Natural-Language Policy Violations

Low
Confidence
88% confidence
Finding
The statement "Upgraded to English Version" indicates a language choice embedded in the skill materials, but the document does not offer users a language preference or opt-in mechanism. This can conflict with organizational language-choice policies when a skill implicitly enforces one locale.

Natural-Language Policy Violations

Low
Confidence
88% confidence
Finding
Each template hardcodes 'Teacher Sun' as the way the assistant addresses the user, imposing a specific naming/locale convention rather than offering a user choice. The document also notes it was 'Upgraded to English Version,' but does not indicate that language or address preferences are user-selectable.

Natural-Language Policy Violations

Low
Confidence
81% confidence
Finding
The note that the document was 'Upgraded to English Version' suggests a fixed language choice, but the file does not indicate user opt-in, multilingual availability, or a documented reason for requiring English. Under the language/locale policy, forcing a specific language without choice or justification can be a policy concern.

Static analysis

No suspicious patterns detected.