T02 · Agent Memory Poisoning
- Location
config/cron.template.json:72- Finding
Unvalidated Conversation Content Is Persisted as Behavioral Memory
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is not clearly malicious, but it needs review because it combines scheduled Feishu sharing, persistent memory updates, and session-history access for sensitive executive work without enough controls.
Before installing, treat this as a Review item: pin the installer and dependencies, use least-privilege Feishu credentials, disable or narrow `session-logs`, and turn off automatic knowledge-base updates until there is explicit review, retention, deletion, and redaction for sensitive executive conversations.
config/cron.template.json:72Unvalidated Conversation Content Is Persisted as Behavioral Memory
config/agents.template.json:31Role Agents Receive Unnecessary Access to Session Logs
README.md:31Installation and Skill Dependencies Are Not Pinned to Audited Versions
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
The routing rule uses a single natural-language keyword, '主脑', as the sole determinant for whether the main agent acts directly or spawns subagents. This is dangerous because ordinary user requests will default into delegated execution, enabling unintended tool use, unnecessary privilege expansion, and easy prompt steering by an attacker who can manipulate phrasing rather than explicit authorization.
The skill hard-codes a specific Chinese phrase as a control-plane trigger without offering equivalent mechanisms for other languages or unambiguous command syntax. This makes behavior easy to spoof or bypass through wording differences and creates inconsistent security boundaries, since execution mode depends on linguistic form instead of authenticated intent.
The skill establishes continuous capture and internalization of corrections, preferences, and repeated instructions into persistent memory and 'lessons'. This is dangerous because attackers can deliberately seed malicious preferences or false operational rules over time, causing durable prompt injection, privacy leakage, and policy drift across future sessions.
The document instructs automatic extraction of user preferences, prohibitions, and instruction patterns into a knowledge base on a schedule. Persisting this behavioral data without clear minimization, consent, retention, or sensitivity filtering can store secrets, sensitive habits, or attack-injected instructions that later influence future behavior.
The README advertises automatic Feishu push notifications, permanent memory, daily auto-optimization, and knowledge-base auto-updates, but does not warn users that business data may be transmitted to external services or modified automatically. In a CEO-oriented decision assistant, this omission is more dangerous because the data involved is likely sensitive, strategic, or regulated.
The installation prerequisites require App Secret and API keys but provide no guidance on secure handling, storage, rotation, or least-privilege use. This increases the chance that users will paste long-lived secrets into plaintext files, screenshots, chats, or version control, leading to credential compromise.
The README instructs users to run npx clawhub@latest install ceo-ai-assistant, which fetches and executes the latest remote package without pinning an exact version. This creates a supply-chain risk: if the upstream package is compromised or a breaking/malicious release is published, users may execute unreviewed code during installation.
The skill advertises Feishu notifications and scheduled briefs but does not clearly warn that potentially sensitive business data may be automatically transmitted to an external messaging platform on a recurring basis. For a CEO-focused assistant handling strategic, financial, and operational decisions, this can lead to unintended disclosure of confidential information if notifications are enabled without explicit data-scope controls and user consent.
The documented memory system stores permanent memories, daily logs, preferences, taboos, and instruction patterns without any retention, minimization, or sensitive-data handling guidance. In a CEO decision-support context, these stores are likely to contain highly sensitive strategic, personnel, financial, and behavioral information, creating long-term exposure if accessed by unauthorized users, reused unexpectedly, or retained indefinitely.
The manifest sets "defaultLanguage": "zh-CN", which indicates the skill is configured to operate in a specific language/locale. Under the policy, locale constraints should either be optional, offer user choice, or be clearly justified as region-specific; this file does not show opt-in or an explicit justification tied to compliance.
The stocks agent enables session-logs, which can capture prompts, responses, and potentially sensitive business or financial information without any visible disclosure or consent mechanism in this configuration. In a CEO decision-support context, logged sessions may include strategic plans, investment reasoning, or confidential company data, increasing privacy and data-governance risk.
The projectmanager agent also enables session-logs without any accompanying notice, which creates a risk of silently retaining operational discussions, internal project details, or confidential business records. Because this skill is designed for executive collaboration and workplace coordination, the context makes undisclosed logging more dangerous than in a low-sensitivity consumer workflow.
Job names and payload messages are written exclusively in Chinese, and the schedules are tied to Asia/Shanghai, which indicates the skill is effectively forcing a specific language/locale. The file does not state that this is a China-specific deployment or offer any user language choice, so it may violate language/locale policy requirements.
The scheduled job explicitly instructs the agent to extract 'new preferences, taboos, and instruction patterns' from daily conversations and update a knowledge base, but the configuration shows no user notice, consent, retention boundary, or review step. This creates a privacy and governance risk because sensitive behavioral data and implicit instructions may be persistently stored without the user's awareness.
The '知识库自动更新' task directs automatic extraction of conversation-derived preferences, prohibitions, and instruction patterns into persistent memory. In a CEO-assistant context, conversations are likely to contain strategic, operational, and personal-sensitive information, so automatic retention materially increases the risk of sensitive data leakage, over-collection, and unintended reuse in later prompts or notifications.
The tagline, description, and shortDescription repeatedly state that the skill is exclusively for '中国CEO' (Chinese CEOs). For a manifest file, this is a natural-language locale/audience restriction, but the text does not present it as an optional preference or explain a compliance-based reason for limiting the skill to that locale.
The README says the assistant is '专为中国CEO设计', which is a locale-specific positioning. Because the document does not explicitly present this as an optional audience choice or provide a compliance-based justification, it may conflict with the policy against forcing a specific language or locale without opt-in.
The description and heading position the skill as specifically for Chinese CEOs, which is a language/locale targeting constraint in natural-language documentation. The file does not indicate this is optional, configurable, or required for a compliance reason, so it may violate the policy against forcing a locale without opt-in.