T02 · Agent Memory Poisoning
- Location
SKILL.md:74- Finding
Conversation-Derived Rules Can Poison Persistent Agent Instructions
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a self-improving memory design, but it asks agents to persist sensitive data and automatically promote learned content into rule files.
Review before installing. Use this only in a controlled workspace, remove any instruction to store API keys or personal identifiers, require explicit approval before writing AGENTS.md or SOUL.md, and add clear retention, deletion, and review controls for any persistent memory.
SKILL.md:74Conversation-Derived Rules Can Poison Persistent Agent Instructions
SKILL.md:125Sensitive Credentials and Personal Identifiers Are Recommended for General-Purpose Persistent Memory
references/consolidation-guide.md:16Untrusted Workspace Documents Can Influence Persistent Routing and Learning State
This section recommends recording sensitive identifiers and environment state including API keys, but does not include a clear warning that such data must not be persisted in agent memory. Because memory systems are often widely queried, distilled, and copied across files, storing these values can turn a convenience feature into a durable secret-exfiltration surface.
The KG schema directly encourages storage of sensitive identifiers, contact information, and API-key-related environment state. In the context of a self-improving persistent memory system, this is especially dangerous because data may be repeatedly indexed, surfaced, copied into summaries, or promoted into other files, amplifying exposure and making eventual leakage more likely.
The trigger conditions are broad enough that the skill may activate for many loosely related requests about memory, optimization, analysis, or automation. Overbroad activation increases the chance that the skill will run in contexts where persistent data collection or automation is not appropriate, causing unintended retention or application of risky memory behaviors.
The file presents the skill instructions entirely in Chinese and includes culturally specific scheduling language, but it does not state that the skill is region-specific or allow the user to choose another language. Under the policy, forcing a language or locale without opt-in is a natural-language policy concern.
The skill instructs storing user profile data such as forms of address, timezone, and preferences without any visible privacy notice, retention policy, or consent mechanism. Even if individually low sensitivity, persistent collection of personal preference and identity-adjacent data can enable profiling and exceed user expectations when embedded in an automated memory system.
The heartbeat workflow directs recurring extraction of conversation memory into persistent stores and periodic consolidation/distillation, creating a broad retention pipeline for user data. Without strict minimization, consent, and deletion controls, this encourages accumulation of personal and contextual information beyond immediate necessity and increases the blast radius of any memory compromise.
The skill explicitly recommends storing highly sensitive identifiers and secrets in the knowledge graph, including OpenID, contact information, and environment state such as API keys. In a generic memory/automation skill, persisting secrets and strong identifiers is unnecessary and materially increases the chance of credential leakage, unauthorized correlation of user identities, and long-term compromise if memory stores are exposed or reused.
The file is entirely written in Chinese and includes Chinese-specific file names and output formats, but it does not offer the user any language/locale choice or explain that the skill is intentionally restricted to a Chinese-language context. Under SQP-3, forcing a specific language without user opt-in is a natural-language policy violation.
The guide documents automatic creation and modification of workspace files, including .access_stats.json, SHADOW.md, and route-table updates, triggered automatically after each heartbeat. In an agent skill context, undocumented autonomous writes can alter persistent memory, create misleading records, and expand the agent’s writable surface without explicit user awareness or consent.
The guide explicitly instructs updating MEMORY.md with inferred user preferences and behaviors, but provides no requirement for user consent, transparency, or confirmation before persisting profile-like data. In a memory-management skill, this is risky because it normalizes silent retention and modification of user-related data, which can lead to privacy violations, inaccurate profiling, or persistence of sensitive information the user did not intend to store.
The monthly promotion workflow directs the agent to write to AGENTS.md, SOUL.md, and errors.md based on internal heuristics, without any authorization, approval, or integrity safeguards. This is more dangerous than ordinary memory updates because AGENTS.md and SOUL.md appear to be governance or rule-bearing files; allowing autonomous modification can permanently alter future agent behavior, create instruction drift, or let poisoned observations become durable policy.
标题及全文内容均以中文编写,且没有说明这是面向特定中文环境的限定文档,也未给出语言可选项。根据规则,若技能在自然语言层面强制特定语言且无用户选择或合理地域性说明,属于语言/locale 策略违规。
A skill document that effectively forces a single language can violate language-choice policy when no opt-in or justification is provided. This file contains all instructions in Chinese and does not mention that the language is optional, configurable, or scoped to a Chinese-speaking audience.