Back to skill

Security audit

starmemo

Security checks for vulnerabilities and agentic risk

Overview

This is a functional memory assistant, but it needs Review because it can automatically retain conversations and send prompts or memories to external services with incomplete disclosure and unsafe setup patterns.

Install only if you are comfortable with a memory skill that may save ordinary conversation text locally and, after LLM configuration, send current prompts or recalled memories to configured AI providers. Avoid entering secrets, use an isolated environment, leave AI/web/key persistence disabled unless needed, and rotate any API key that may have been passed via command-line history.

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 (4)

T09 · Insecure Skill Coding Practices

Error
Location
starmemo.py:319
Finding

Automatic Collection and External Transmission of Conversation Data

Content
View full analysis
str: if not self.enable_ai: return "" headers = { "Content-Type": "application/json", "Authorization": f"Bearer {self.api_key}" } data = { "model": self.model, "messages": [{"role": "user", "content": prompt}], "temperature": self.temperature, "max_tokens": max_tokens } try: res = requests.post( self.api_endpoint, headers=headers, json=data, timeout=15 ) res.raise_for_status() return res.json()["choices"][0]["message"]["content"].strip() except Exception as e: print(f"⚠️ LLM call failed: {e}") return "" ``` ```python structured = self.ai.extract_structured(text) self.storage.save_daily( cause=structured.get("cause", text[:100]), change=structured.get("change", text[:200]), todo=structured.get("todo", ""), topic=structured.get("topic", "") ) knowledge_list = self.ai.extract_knowledge(text) ``` ### Technical Analysis The legacy input hook defaults to saving every user input when the `save` configuration field is absent. In the legacy message handler, almost every non-configuration message is passed to `LLMClient.optimize()` before be ...[truncated 2551 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
starmemo.py:89
Finding

Configurable Endpoint Can Receive API Credentials and Private Memory Data

Content
View full analysis
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
starmemo.py:40
Finding

Import-Time Installation of an Unpinned Dependency

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
cli.py:62
Finding

API Key Exposure Through Command-Line Arguments and Configuration Output

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Findings (48)

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared purpose presents the skill as a memory system, but the detected behavior includes broad automatic data capture, third-party LLM transmission, web fetching, persistent config/API key storage, and package installation. This mismatch is highly dangerous because users and reviewers may consent to a simple memory feature while the skill actually exfiltrates data, expands its runtime surface, and stores sensitive material.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

All user input is automatically persisted by default without a user-facing warning or consent flow. In a memory skill, this is especially dangerous because users may disclose secrets, credentials, personal data, or regulated information that will then be stored locally across sessions.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The code explicitly states and implements global capture of all user inputs across platforms, with saving enabled by default and no scenario-based exceptions. In context, that makes the skill more dangerous because it is integrated into chat/agent workflows where sensitive data is commonly entered unexpectedly.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
93% confidence
Finding

The comments claim the original retrieval/response flow remains unchanged, but the referenced helper name does not exist for query processing and instead collides with a save function name elsewhere. In security-sensitive review, misleading comments plus incompatible control flow are dangerous because they can hide unintended behavior, break expected safeguards, or cause operators to misunderstand what data is stored or processed.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The reusable save function persists arbitrary provided content to disk without contextual checks, policy enforcement, or sensitivity filtering. Because it is presented as cross-platform and unrestricted, other components can call it to store sensitive material broadly and silently.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

Automatic dependency installation is not necessary for the core stated purpose and gives the skill the ability to modify the host environment. In a plugin setting, that is dangerous because loading the skill can trigger network access and package execution without a review gate.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill sends user input and memory context to third-party LLM APIs without an in-band warning at the time of use. Because recalled memory can contain prior sensitive conversations, this creates a meaningful data exfiltration risk beyond the current prompt alone.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The README advertises web-enabled learning and external LLM integration via API keys, but does not clearly disclose that user memories, prompts, or knowledge-base content may be transmitted to third-party services. In a memory skill, that omission is especially risky because the stored data is likely to contain sensitive personal or organizational information, creating privacy, compliance, and unintended data-sharing exposure.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
87% confidence
Finding

The skill advertises capabilities that imply file access, persistence, shell execution, and network use, but it does not declare any explicit tool scope or permissions boundary. In a memory-oriented skill, this is dangerous because it obscures the real attack surface and can enable over-privileged operation, especially when combined with persistence and external calls.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill explicitly enables online learning and persistence but does not warn users that their data may be stored long-term or transmitted to external services. In this context, the omission is dangerous because a memory system naturally handles user-provided content that may include sensitive personal or business information.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The save triggers are broad enough to capture ordinary conversation, task details, and newly learned information without a narrowly defined consent boundary. In a persistent memory skill, that creates a substantial risk of storing sensitive personal, enterprise, or cross-context data that the user did not intend to retain.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

Using common conversational words like '之前', '上次', or '那个' as recall triggers is too vague and can cause memory retrieval in contexts where the user is not actually asking for persisted data. This increases the chance of unintended disclosure of previously stored sensitive information into the current conversation.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill registers the broad on_user_input hook, meaning it can activate on every user message without any declared trigger constraints or exclusions. For a skill with read_file and write_file permissions and a memory-oriented purpose, this increases the attack surface for unintended data collection, silent persistence of sensitive content, and unexpected file modifications.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
81% confidence
Finding

The save command reads arbitrary content, including from stdin, and writes it to storage via skill.storage.save. While it prints a success message afterward, there is no prior warning in the command help, docstring, or inline user disclosure that provided content will be persisted as memory, which may surprise users handling sensitive text.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

Displaying part of a stored API key without warning increases the chance of accidental credential exposure through screenshots, shell history capture, terminal logging, or shared sessions. In a CLI context, users often assume 'show config' is safe to run, so exposing secret material is riskier than necessary.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

Automatic package installation occurs without user/admin warning, which means simply importing the skill may trigger environment changes and outbound network activity. This violates least surprise and can undermine controlled deployment practices.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
96% confidence
Finding

The skill automatically invokes pip at import time to install a dependency, which executes a package-management side effect without user approval. In an agent/plugin context this expands the trusted computing base, can alter the runtime environment, and may introduce supply-chain risk if package sources or mirrors are compromised.

Content

Scanner excerpt · starmemo.py (reported line 45)May include surrounding context.

python
import requests
    except ImportError:
        import subprocess
        subprocess.check_call([sys.executable, "-m", "pip", "install", "-q", "requests"])
auto_install()
import requests

External Transmission

Medium
Category
Data Exfiltration
Confidence
97% confidence
Finding

This request transmits user-provided text to an external LLM provider for optimization, potentially including sensitive information from raw inputs. Because it is tied to memory processing, the function may exfiltrate data users believed would remain local.

Content

Scanner excerpt · starmemo.py (reported line 164)May include surrounding context.

python
"max_tokens": self.config.optimize_tokens
        }
        try:
            res = requests.post(self.api, headers=headers, json=data, timeout=10)
            res.raise_for_status()
            return res.json()["choices"][0]["message"]["content"].strip()[:self.config.max_len]
        except:

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

This memory skill performs broader network-backed chat and optional retrieval functions rather than staying limited to local memory management. That capability expansion increases the attack surface and privacy risk because user inputs and recalled memory may be transmitted externally in ways users do not expect from a 'memory' component.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
98% confidence
Finding

This call sends current user input and memory context to an external LLM chat endpoint. The combination of live prompt data plus stored memory materially raises confidentiality risk if the memory contains prior sensitive content.

Content

Scanner excerpt · starmemo.py (reported line 190)May include surrounding context.

python
"max_tokens": self.config.chat_tokens
        }
        try:
            res = requests.post(self.api, headers=headers, json=data, timeout=10)
            res.raise_for_status()
            return res.json()["choices"][0]["message"]["content"].strip()
        except:

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

Fallback and system prompt strings in clarify are hard-coded in Chinese, and the file branding/instructions are also Chinese-centric. This can violate language/locale policy because users are not offered a language choice or an opt-in to Chinese-only responses.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
96% confidence
Finding

This clarify flow also sends user input and memory context to an external endpoint, even though the task could be handled locally or with a minimal template. It creates unnecessary external exposure for a non-essential feature.

Content

Scanner excerpt · starmemo.py (reported line 209)May include surrounding context.

python
]
        data = {"model": self.config.model_name, "messages": messages, "temperature": self.config.chat_temperature, "max_tokens": self.config.clarify_tokens}
        try:
            res = requests.post(self.api, headers=headers, json=data, timeout=10)
            res.raise_for_status()
            return res.json()["choices"][0]["message"]["content"].strip()
        except:

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

Optional web fetching is outside the minimum capability needed for structured memory and causes user-derived queries to be sent to external services. Even if gated by config, it broadens the skill from storage into retrieval/augmentation and increases privacy and exfiltration risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

When enabled, the skill sends search queries derived from user messages to external web services without a user-facing warning at request time. Queries may contain names, internal project details, or other sensitive content that users did not intend to disclose externally.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill explicitly advertises persistent memory storage and web-connected learning, and later shows both features enabled, but provides no explicit consent flow, retention notice, or boundaries on what data may be transmitted externally. In a memory-oriented skill, this materially increases the chance that user-provided sensitive information is stored long-term and potentially sent to third-party LLM or web services without clear user awareness.

Content

No source excerpt is available for this finding.