T02 · Agent Memory Poisoning
- Location
scripts/learner.py:92- Finding
Unvalidated user preferences are persisted and intended for automatic reuse
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is mostly a faceless-video workflow, but it also adds persistent self-learning behavior that stores user data and can influence future skill instructions.
Review before installing. Use this only if you are comfortable with it recording usage, errors, and preferences into a local JSON file and potentially shaping later behavior from that data. Do not put credentials, private business information, personal data, or instruction-like text into its notes or preferences; consider disabling or removing the learner before use.
scripts/learner.py:92Unvalidated user preferences are persisted and intended for automatic reuse
scripts/learner.py:58Free-form notes and preferences are retained in plaintext without access hardening
SKILL.md:213Skill instructions permit persistent self-modification based on attacker-controlled learning data
The skill is declared as a faceless-video automation workflow, but it also introduces a generic cross-session learning and persistence subsystem unrelated to that purpose. That scope expansion is dangerous because it can silently collect usage history, error data, and user preferences beyond what users would reasonably expect from a video-production skill, increasing privacy and misuse risk.
The listed trigger words include generic phrases like “AI视频” and “频道自动化,” which are broad enough to match many ordinary content-creation requests beyond this specific skill. The manifest does not provide scope limits, activation constraints, or negative examples to clarify when the skill should not activate.
The documented self-evolution module persists cross-session usage history and user preferences, which exceeds what is necessary for generating faceless videos. This creates a data-retention surface where behavioral data can accumulate over time without a narrowly scoped operational need, making accidental overcollection or later misuse more likely.
The skill documentation describes automatic writing of learned preferences and usage history without clearly warning users in the main skill description or obtaining consent. This is dangerous because users may share information under the assumption of a one-off content-generation workflow, while the skill quietly converts it into retained profile data.
Persistent storage of user preferences, error notes, and usage history in a natural-language JSON file creates a clear data-retention risk. Free-form notes can easily capture sensitive or identifying details, and because the skill is public and the retention is open-ended, the context makes this more dangerous than ordinary transient workflow state.
The learning-module examples reference unrelated capabilities such as resume optimization and price comparison, showing that the persistence layer is generic rather than confined to video automation. That broadens the effective capability surface and suggests the skill could be repurposed to store or reason about unrelated user tasks and preferences, which is inconsistent with least privilege.
The manifest describes an end-to-end faceless video production workflow for platforms like Bilibili, YouTube, and Douyin. This file instead implements a reusable, skill-agnostic telemetry and preference-learning system for any WorkBuddy skill, which is a materially different capability not described as part of the video-automation workflow.
This code file contains natural-language documentation and CLI output exclusively in Chinese, including the module description and usage instructions. That imposes a locale/language choice on users without offering an alternative or opt-in, which matches the language-policy violation criteria.
The module persistently stores operational history, error notes, and arbitrary user preferences for any supplied skill directory, including under ~/.workbuddy/skills, without any access control, consent flow, retention policy, or data minimization. In a skill ecosystem, this creates a privacy and cross-skill data collection risk because free-form notes and preferences may capture sensitive user or workflow information and remain on disk indefinitely.
The example preference-setting command hardcodes 输出语言 --val 中文, which indicates a default Chinese-language preference without stating that users may choose another language. This can conflict with language/locale choice expectations when the skill is presented publicly and also includes English metadata.