Evermemos

• Outcome: Review
• Latest audit: Sep 12, 2026 · 2:17 AM UTC
• Version: 1.0.0

Overview

This memory skill is purpose-aligned, but it can automatically send and persist conversation and profile data to a configurable EverMemOS server without clear user-control, retention, HTTPS, or authentication safeguards.

Install only if you are comfortable with an agent storing selected conversation content, preferences, and profile details in an EverMemOS server. Use a trusted HTTPS endpoint, require API authentication, disable or tightly review automatic memory capture, avoid storing secrets or regulated data, and ensure users can inspect and delete stored memories.

A.I.G

Vulnerability Patterns

• Insecure Skill Coding Practices: Finds exploitable flaws such as hardcoded secrets or command injection
• Skill Instruction Hijacking: Alters the agent's session goals or safety constraints when the skill loads
• Agent Memory Poisoning: Writes attacker-controlled rules into memory that affect later sessions
• Remote Payload Retrieval and Execution: Fetches external code whose behavior can change after review
• Embedded Malicious Code: Ships malicious scripts inside the skill and executes them locally

Findings (2)

T09 · Insecure Skill Coding Practices

Warning

• Location: SKILL.md:79
• Finding: Automatic Transmission of Potentially Sensitive Conversation Data<![CDATA[ ## Vulnerability Details **File Location**: `SKILL.md:79-89`, `SKILL.md:160-165`, and `SKILL.md:177-211` **Vulnerability Type**: Excessive automatic collection and network transmission of user data **Risk Level**: Medium ### Vulnerable Code ```python def save_conversation_summary(messages): """从对话中提取关键点并存储""" for msg in messages: if is_important(msg): # 判断是否为关键信息 store_memory( content=msg["content"], sender=msg["sender"], metadata={"type": "conversation_summary"} ) ``` The automatic storage policy is documented as follows: ```text 在以下时机自动存储记忆： 1. **对话结束** - 提取关键要点 2. **用户自我介绍** - 存储用户信息 3. **任务完成** - 记录完成内容 4. **用户偏好表达** - 记住偏好设置 ``` The complete example sends the selected content to the configured server: ```python class EverMemOS: def __init__(self, url=None, user_id="default"): self.base_url = url or os.getenv("EVERMEMOS_URL", "http://localhost:1995") self.user_id = user_id def store(self, content, sender="user"): """存储记忆""" return requests.post( f"{self.base_url}/api/v1/memories", json={ "message_id": f"msg_{int(time.time()*1000)}", "content": content, "sender": sender, "user_id": self.user_id, "create_time": datetime.utcnow().isoformat() + "Z", "scene": "assistant" } ).json() ``` ### Technical Analysis Network-based storage is intrinsic to the declared long-term memory functionality. However, the Skill directs the agent to store conversation summaries, introductions, task records, and preferences automatically. These categories can contain personally identifiable information, credentials accidentally pasted into a conversation, confidential business information, or other sensitive material. The documented workflow provides no explicit per-item consent gate, content pre ...[truncated 1672 chars]
• Remediation: <![CDATA[ ## Remediation Suggestions 1. Disable automatic memory storage by default. 2. Require explicit user confirmation before each storage operation. 3. Show the exact content, user identifier, and destination that will receive the data. 4. Store only the minimum fact requested by the user instead of full messages or broad summaries. 5. Add a sensitive-data filter that removes credentials, authentication tokens, financial information, health information, and unnecessary personal identifiers. 6. Restrict destinations to an administrator-approved allowlist of trusted origins. 7. Separate memory namespaces by authenticated user identity and prevent use of a shared default identity in multi-user deployments. 8. Define retention periods and provide list, correction, export, and deletion controls. 9. Record auditable consent and storage events without duplicating sensitive content in logs. 10. Treat retrieved memories as untrusted data and prevent stored content from becoming agent instructions. ]]>

T09 · Insecure Skill Coding Practices

Warning

• Location: SKILL.md:60
• Finding: Sensitive Memory Operations Do Not Enforce Secure Transport or Authentication<![CDATA[ ## Vulnerability Details **File Location**: `SKILL.md:20-24`, `SKILL.md:60-72`, and `SKILL.md:177-211` **Vulnerability Type**: Plaintext-capable transmission and omitted API authentication **Risk Level**: Medium ### Vulnerable Code The configuration documents an API key and a plaintext default URL: ```text 在环境中设置以下变量： - `EVERMEMOS_URL` - EverMemOS API 地址 (默认: http://localhost:1995) - `EVERMEMOS_API_KEY` - API Key (如需要) ``` The storage implementation neither validates the destination scheme nor applies the documented API key: ```python import os import requests from datetime import datetime EVERMEMOS_URL = os.getenv("EVERMEMOS_URL", "http://localhost:1995") def store_memory(content, sender="user", user_id="default"): """存储记忆到 EverMemOS""" data = { "message_id": f"msg_{int(datetime.now().timestamp()*1000)}", "content": content, "sender": sender, "user_id": user_id, "create_time": datetime.utcnow().isoformat() + "Z", "scene": "assistant" } resp = requests.post(f"{EVERMEMOS_URL}/api/v1/memories", json=data) return resp.json() ``` The complete client example has the same issue for storage and retrieval: ```python class EverMemOS: def __init__(self, url=None, user_id="default"): self.base_url = url or os.getenv("EVERMEMOS_URL", "http://localhost:1995") self.user_id = user_id def store(self, content, sender="user"): """存储记忆""" return requests.post( f"{self.base_url}/api/v1/memories", json={ "message_id": f"msg_{int(time.time()*1000)}", "content": content, "sender": sender, "user_id": self.user_id, "create_time": datetime.utcnow().isoformat() + "Z", "scene": "assistant" } ).json() def recall(self, query, top_k=5): """检索记忆""" return requests.post( f"{self.base_url}/ap ...[truncated 2662 chars]
• Remediation: <![CDATA[ ## Remediation Suggestions 1. Require HTTPS for every non-loopback EverMemOS endpoint. 2. Permit plaintext HTTP only for explicit development mode and verified loopback addresses such as `127.0.0.1`, `::1`, or `localhost`. 3. Reject URLs with unexpected schemes, embedded credentials, fragments, or unapproved hosts. 4. Read `EVERMEMOS_API_KEY` and send it using the server's supported authorization header, such as `Authorization: Bearer <token>`. 5. Fail closed when authentication is required but the key is absent. 6. Never place API keys in query strings or log them. 7. Configure connection and response timeouts for all requests. 8. Check HTTP status codes with `raise_for_status()` and handle invalid JSON safely. 9. Use trusted certificate verification and do not recommend disabling TLS validation. 10. Document server-side authorization, user isolation, key rotation, revocation, and rate limiting. 11. Add tests proving that remote HTTP URLs are rejected and authenticated HTTPS requests include the expected header. ]]>

SkillSpector

Vulnerability Patterns

• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Trigger Abuse: Overly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access
• Supply Chain: Unpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (9)

Missing User Warnings

High

• Confidence: 98% confidence
• Finding: The skill presents itself as a production memory integration but does not warn users that conversation content may be stored remotely on an EverMemOS server. In a memory skill, failure to disclose remote persistence and transmission undermines informed consent and creates substantial privacy and compliance risk.

Missing User Warnings

High

• Confidence: 98% confidence
• Finding: The document says the system will automatically record user profile information and preferences, but it gives no warning about the data impact of collecting this personal information. In this context, profiling and long-term storage are especially sensitive because they can accumulate behavioral and preference data over time without meaningful user awareness.

External Transmission

Medium

• Category: Data Exfiltration

Content

### 1. 基本存储 (单条消息)

```bash
curl -X POST ${EVERMEMOS_URL}/api/v1/memories \
  -H "Content-Type: application/json" \
  -d '{
    "message_id": "msg_001",

• Confidence: 90% confidence
• Finding: The curl example demonstrates sending message content directly to the EverMemOS API, including user-authored text and metadata. As documentation for a memory skill this is functional, but it normalizes remote transfer of conversational data without emphasizing secure transport, authentication, or user consent, which increases privacy risk in real deployments.

External Transmission

Medium

• Category: Data Exfiltration

Content

"create_time": datetime.utcnow().isoformat() + "Z",
        "scene": "assistant"
    }
    resp = requests.post(f"{EVERMEMOS_URL}/api/v1/memories", json=data)
    return resp.json()

# 使用示例

• Confidence: 92% confidence
• Finding: This code transmits memory content to an externally configured EverMemOS endpoint using requests.post, which can include user conversation data and identifiers. External transmission is expected for this skill, but it remains security-relevant because the endpoint is configurable, the default uses plain HTTP, and the example does not show authentication, consent checks, or data minimization.

Vague Triggers

Medium

• Confidence: 95% confidence
• Finding: The trigger phrases include broad everyday expressions such as “之前说过...”, “保存这个”, and “我的偏好是...”, which can cause the memory skill to activate unexpectedly during normal conversation. In this skill’s context, unintended activation is risky because it can lead to storage or retrieval of user data without a clear, explicit request.

Vague Triggers

Medium

• Confidence: 97% confidence
• Finding: The automatic memory triggers are vague and include sensitive situations like self-introduction, task completion, and preference expression without clear limits on what may be stored. Because this skill sends conversation content to a long-term memory system, ambiguous auto-save rules can cause overcollection of personal data and retention of information the user did not intend to persist.

External Transmission

Medium

• Category: Data Exfiltration

Content

def store(self, content, sender="user"):
        """存储记忆"""
        return requests.post(
            f"{self.base_url}/api/v1/memories",
            json={
                "message_id": f"msg_{int(time.time()*1000)}",

• Confidence: 93% confidence
• Finding: The store method sends conversation content and user identifiers to the configured memory server. In a long-term memory skill this is expected behavior, but it is still dangerous if used without strong endpoint validation, encryption, and clear user authorization because it can persist sensitive conversation data outside the immediate chat environment.

External Transmission

Medium

• Category: Data Exfiltration

Content

def recall(self, query, top_k=5):
        """检索记忆"""
        return requests.post(
            f"{self.base_url}/api/v1/memories/retrieve",
            json={
                "query": query,

• Confidence: 86% confidence
• Finding: The recall method sends the user's query and identifier to the remote retrieval endpoint, exposing potentially sensitive search terms and linking them to a persistent profile. Although retrieval is core functionality, transmitting such queries externally without explicit disclosure and transport safeguards can leak sensitive interests or prior conversation context.

Natural-Language Policy Violations

Low

• Confidence: 84% confidence
• Finding: 文件中的名称、描述和触发短语均以中文固定表达呈现，尤其触发条件仅列出中文短语，没有说明是否支持其他语言或允许用户选择语言。若组织要求避免未经选择地强制特定语言/地区，这种默认中文约束可能构成语言策略问题。

Static analysis

No suspicious patterns detected.