T02 · Agent Memory Poisoning
- Location
SKILL.md:221- Finding
Uncontrolled Retention of Customer Feedback and Persistent Knowledge Mutation
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 221–225
Vulnerability Type: Persistent memory poisoning and unspecified customer-data retention
Risk Level: MediumComplete instruction excerpt, translated into English from the source:
markdown After each risk-control task: 1. Record customer feedback (whether the risk assessment was accurate) 2. Update the risk-feature library (new default patterns) 3. Update the knowledge base (new regulatory policies) 4. Mark ambiguous cases (for subsequent training)Technical Analysis
The Skill instructs the Agent to record customer feedback and modify persistent risk and knowledge repositories after every completed task. The instructions do not define:
- Whether explicit user consent is required
- Which submitted data may be retained
- How confidential contract or counterparty information must be sanitized
- How feedback and regulatory claims must be validated
- Whether updates are isolated from production knowledge
- Access-control, retention, review, rollback, or deletion procedures
Contract reviews and counterparty investigations can contain confidential commercial terms, personal information, financial records, and internal approval data. Automatically retaining such material creates a data-governance risk. Moreover, treating customer feedback or submitted cases as knowledge-update material creates a memory-poisoning channel: attacker-controlled or inaccurate claims could be incorporated into persistent state and affect later users.
The supplied Python scripts do not implement a persistent memory store, so exploitation depends on whether the hosting Agent provides memory or knowledge-base mutation capabilities. No operating-system persistence, remote communication, credential access, or arbitrary code execution was identified.
Attack Path
- An attacker submits a crafted contract, fabricated regulatory stat ...[truncated 1309 chars]
- Remediation
View remediation
Remediation Suggestions
- Remove the unconditional instruction to record information after every task.
- Require explicit, informed user consent before retaining feedback or case material.
- Default to session-only processing and prohibit persistence of raw contracts, personal information, financial records, and counterparty documents.
- Permit knowledge updates only through a separately authorized administrative workflow.
- Validate proposed regulatory and risk-feature updates against authoritative sources before publication.
- Store proposed updates in a quarantined review queue rather than directly modifying production knowledge.
- Apply data minimization and redaction so that only sanitized, structured risk indicators can be retained.
- Define access controls, tenant isolation, retention periods, deletion procedures, audit logging, versioning, and rollback controls.
- Add an explicit instruction that user feedback and uploaded documents are untrusted data and must never become persistent rules automatically.
- Replace the affected block with wording such as:
markdown Do not retain customer content or update persistent memory automatically. Feedback may be summarized only with explicit user consent and after removing confidential or personal information. Knowledge-base changes require review against authoritative sources and approval by an authorized administrator.
