T02 · Agent Memory Poisoning
- Location
SKILL.md:105- Finding
Autonomous persistence of conversation content without explicit user consent
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a disclosed shared-memory tool, but it can silently store and later retrieve personal cross-agent memories without clear user consent or isolation.
Install only if you are comfortable with a shared, persistent memory database that agents may update and search across personas. Before using it, disable or avoid automatic writes, require explicit confirmation before storing anything, keep sensitive health/family/work details out of it unless intended, and store the KB in a private local directory with restrictive permissions.
SKILL.md:105Autonomous persistence of conversation content without explicit user consent
scripts/memory_kb.py:34Sensitive memory data is stored in plaintext without enforced restrictive permissions
scripts/memory_kb.py:123Missing file locking and non-atomic index updates permit concurrent corruption
The skill is explicitly designed for cross-identity shared memory but does not require a clear privacy warning or consent flow before persisting user content. Because data can be reused by multiple agent personas later, users may unknowingly disclose sensitive information far beyond the current conversation.
The skill tells the agent to trigger based on broad semantic judgment rather than a clear user command, so ordinary conversation can be interpreted as authorization to persist or retrieve memory. In this context, the broad trigger is especially dangerous because the action affects long-lived shared storage across agent identities.
Proactive storage based on semantic cues enables silent collection of user content, including potentially sensitive reflections, decisions, or experiences. Because the capture is driven by heuristic interpretation rather than explicit direction, the agent may over-collect and persist information the user never intended to save.
The instructions allow automatic memory writes without prior confirmation, which enables silent persistence of user-derived content. In a shared memory system, this creates substantial privacy risk because sensitive or contextual statements may be stored permanently without informed consent.
The notes explicitly direct the agent to store or retrieve data without explicit user request and to confirm storage silently, while omitting any warning about persistence or cross-session reuse. This normalizes covert collection and retrieval behavior in a context involving personal knowledge retention.
Instructing the agent to invoke storage or retrieval without explicit user request raises the risk of unintended disclosure across sessions and personas. Retrieval is particularly sensitive because it can surface prior information in a new context, exposing data that the user did not mean to reuse.
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
删除配置文件即可恢复默认值:
rm ~/.openclaw/memory/config.json
The skill declares filesystem and environment-dependent behavior but does not define any explicit tool scope or permissions boundary. That makes its effective capabilities opaque to reviewers and increases the chance an agent can read/write persistent data or use environment-derived paths beyond what users expect.
Automatic cross-persona memory capture can retain sensitive user information and later surface it in a different persona or session where the user did not expect disclosure. The cross-identity design materially increases the danger because information collected in one context becomes available in others.
This duplicate finding points to the same persistent storage setup in the dependency block. In context, the risk comes from durable, shared retention of potentially sensitive user content rather than the mkdir command alone.
description: 为 OpenClaw 用户所有 Agent 身份提供统一的跨身份共享记忆层,支持写入、检索、浏览、关联和回顾个人知识库;当用户需要记录洞察、检索经验、总结知识或定期回顾时使用
dependency:
system:
- mkdir -p ~/.openclaw/memory/index
---
# Shared Memory Knowledge Base
This duplicate finding points to the same persistent storage setup in the dependency block. In context, the risk comes from durable, shared retention of potentially sensitive user content rather than the mkdir command alone.
description: 为 OpenClaw 用户所有 Agent 身份提供统一的跨身份共享记忆层,支持写入、检索、浏览、关联和回顾个人知识库;当用户需要记录洞察、检索经验、总结知识或定期回顾时使用
dependency:
system:
- mkdir -p ~/.openclaw/memory/index
---
# Shared Memory Knowledge Base
The activation scope overlaps with common requests like summarizing, reflecting, or recalling prior experience, which are normal assistant tasks and not necessarily requests to access a persistent cross-session knowledge base. This ambiguity can cause unintended invocation and surprise data handling.
Silent confirmation after writing memory hides that persistence occurred, reducing user awareness and undermining meaningful consent. This makes covert retention easier to normalize and harder for users to audit or correct.
This file presents all operational guidance and warnings in Chinese, effectively imposing a specific language for the skill documentation. The policy allows locale-specific constraints only when documented and justified, or when users are offered a language choice, neither of which is present here.
The document explicitly describes automatic memory writes and automatic importance adjustments that can modify a user's knowledge base, but it does not prominently warn that user data may be created or altered automatically. In a shared-memory skill spanning multiple agent identities, silent or poorly disclosed automatic modification increases the risk of unintended persistence, privacy issues, and hard-to-audit data changes.
This markdown file presents the skill documentation almost entirely in Chinese, including headings, field descriptions, and operational guidance. Under the language/locale policy, forcing a specific language without user opt-in or a documented regional justification is a natural-language policy concern.
This markdown file is written almost entirely in Chinese and presents the taxonomy as the expected format without stating that other languages or locales are supported. Under the policy rule, language constraints should either offer user choice or be clearly documented as a justified region-specific requirement.
This code file contains natural-language descriptions entirely in Chinese, including the top-level module description and operation labels, with no indication that users may choose another language. Under the policy, forcing a specific language without user opt-in is a natural-language locale violation unless the constraint is explicitly justified, which is not present here.
The skill is explicitly designed as a cross-persona shared memory layer and returns prior stored entries via query, list, similar-memory matching, and reflection without any access-control, minimization, or persona isolation enforcement. In an agent environment, this can expose sensitive data recorded by one persona or workflow to another later interaction, creating a confidentiality breach and enabling inadvertent prompt/context leakage across identities.
The script exposes administrative mutation capabilities (rebuild-index, recalc-importance) beyond the stated shared-memory user functions. In an agent-skill context, undocumented maintenance actions increase attack surface and can alter or rewrite knowledge-base state in ways the user may not expect, especially because recalc_importance performs bulk updates and rebuild_index rewrites metadata/indexes.
The operational instructions, triggers, and examples are presented only in Chinese, which implicitly forces a specific language experience. The file does not indicate that users may choose another language or that the Chinese-only constraint is intentional and justified for a specific region or audience.
The reset instructions tell users to delete the configuration file with rm but do not clearly label this as a destructive action or explain the consequences. Even though only configuration is affected, undocumented deletion guidance can lead to accidental loss of custom settings and unsafe copy-paste behavior.
_update_entry is documented as '更新记忆条目(创建新版本)', which implies preserving the existing entry and then creating a new version. Instead, it first modifies the passed-in entry object by setting deleted=True and writing it out, then copies that mutated object and removes the deleted field, so the old object is not preserved as-is but actively transformed into a tombstone record.