T02 · Agent Memory Poisoning
- Location
main.py:47- Finding
Persistent Agent Memory Poisoning Through Untrusted Interaction Details
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is locally oriented and not overtly malicious, but it asks agents to persist broad raw interaction and environmental memory automatically without clear consent, retention, deletion, or isolation controls.
Install only if you intentionally want a local persistent-memory experiment and are prepared to control what gets recorded. Do not connect it to private chats, sensors, local agent memory retrieval, or other agents unless you add explicit opt-in, review, redaction, retention limits, deletion controls, and safeguards that treat stored memories as untrusted data.
main.py:47Persistent Agent Memory Poisoning Through Untrusted Interaction Details
The primary natural-language description is written entirely in Chinese and presents the skill identity and operating model in that language without offering any language or locale choice. Under the policy criteria, forcing a specific language without user opt-in is a natural-language policy concern unless the locale restriction is explicitly justified.
The function appends arbitrary interaction details to a persistent local 'hippocampus' log, encouraging broad retention of raw activity data. Because details is written as-is and tags defaults to a mutable list, the design promotes indiscriminate collection and storage of potentially sensitive information, creating privacy and data-handling risk even without a network exfiltration path.
This routine permanently consolidates previously logged interaction details into persistent markdown files without any minimization, consent, retention limit, or sensitivity filtering. If the cached logs contain user prompts, secrets, personal data, or operational details, the skill creates durable local artifacts that increase exposure and make accidental disclosure or later compromise more damaging.
The manifest sets the skill language to "zh-CN", which indicates a fixed language/locale choice. Under the policy, locale constraints should either offer user choice or be clearly documented and justified as region-specific; this file provides neither.
The skill explicitly directs the agent to record physical environment changes and user preference updates into memory, but it provides no user notice, consent flow, retention limits, or access controls. This creates a real privacy risk because sensitive behavioral and contextual data may be persistently collected and later reused without the user's awareness.
The document explicitly proposes a background tool that silently intercepts raw interaction logs and multimodal sensory input, with no consent, notice, minimization, or access-control language. In an agent skill context, this creates covert surveillance and broad collection of potentially sensitive user data, including conversations and environmental signals, which could be exfiltrated, misused, or retained indefinitely.
The file instructs persistent capture of raw user interactions and sensory inputs, then promotion of processed content into long-term memory, creating a durable memory-capture surface. In a skill for autonomous agents, this is especially dangerous because it normalizes broad surveillance and retention of conversational, contextual, and possibly physical-environment data without clear safety boundaries.
The whitepaper mandates persistent long-term memory storage as a core design feature but does not define retention boundaries, user consent, deletion rights, or sensitivity filtering. Long-lived memory files materially increase the blast radius of compromise by preserving personal data, behavioral patterns, and environmental context beyond the immediate session.
The document describes cross-agent memory sharing over a local network, restricted only by proximity/identity concepts, but provides no warning, consent model, authentication detail, or data-sharing safeguards. Even on a local network, peer-to-peer exchange of memory artifacts can leak personal or operational data across agents and expands the trust boundary in ways users may not expect.
This markdown file presents its primary content in Chinese, including headings and release notes, without indicating that language selection is optional or that the skill is intentionally limited to a Chinese-speaking context. Under the policy rule for natural-language violations, forcing a specific language without user opt-in can be a locale-policy issue.
The natural-language instructions throughout the skill are entirely in Chinese, with no indication that users may choose another language or locale. Under the policy rule, forcing a specific language without opt-in can be a natural-language policy concern unless the locale restriction is documented and justified.