T02 · Agent Memory Poisoning
Error
- Location
index.js:35- Finding
Untrusted Daily Log Content Can Poison Persistent Agent Memory
- Content
View full analysis
- Remediation
View remediation
Security audit
Memory Enhancer
Security checks for vulnerabilities and agentic risk
Overview
This skill has a coherent memory-enhancement purpose, but it automatically sends personal daily logs to an external model and lets a spawned agent write long-term memory without review.
Review this before installing if your daily logs contain private, sensitive, regulated, or third-party information. Enable it only if you are comfortable with automatic external model analysis and model-written long-term memory entries, and prefer a version with explicit opt-in, preview-before-write, schema validation, and a local-only or redacted processing option.
Vulnerability Patterns
- Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
- Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
- Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
- Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
- Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (1)
Vulnerability Patterns
- Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
- Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
- Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
- Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
- Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (6)
Ssd 3
High
- Category
- Not specified by scanner
- Confidence
- 97% confidence
- Finding
The skill is designed to extract personal details from daily logs and persist them into long-term memory stores such as preferences, contacts, habits, and projects. In this context, the absence of privacy controls, validation, retention limits, and injection-resistant handling makes the behavior dangerous because sensitive information is automatically transformed into durable profile data that may be inaccurate, overbroad, or unauthorized.
- Content
Missing User Warnings
Medium
- Category
- Not specified by scanner
- Confidence
- 96% confidence
- Finding
The skill explicitly states that a sub-agent sends daily log content to OpenRouter for analysis, but it does not warn users that potentially sensitive diary or memory data will be transmitted to an external third-party service. Because the skill runs automatically via cron and processes personal logs, users may unknowingly expose private information, contacts, habits, and project details outside the local environment.
- Content
Context-Inappropriate Capability
Medium
- Category
- Not specified by scanner
- Confidence
- 95% confidence
- Finding
The prompt explicitly instructs the sub-agent to modify several persistent memory files and append to a top-level MEMORY.md file based on free-form log content. Because these writes are driven by model interpretation rather than deterministic code, the skill allows long-term state changes from potentially adversarial or sensitive input without validation, review, or tight scope control.
- Content
Missing User Warnings
Medium
- Category
- Not specified by scanner
- Confidence
- 92% confidence
- Finding
The skill instructs a sub-agent to update multiple persistent files and append to MEMORY.md, but there is no user-facing notice that these writes will occur. Silent persistence of interpreted personal data increases the risk of unwanted retention, inaccurate memory entries, and user surprise about the scope of modifications.
- Content
Missing User Warnings
Medium
- Category
- Not specified by scanner
- Confidence
- 96% confidence
- Finding
The skill sends the full contents of a daily memory log to an external model endpoint without any disclosure, minimization, or consent mechanism. Since the logs may contain personal contacts, habits, projects, and other sensitive details, this creates a privacy and data-governance exposure even if the model behaves as intended.
- Content
Context-Inappropriate Capability
Medium
- Category
- Not specified by scanner
- Confidence
- 91% confidence
- Finding
The skill delegates a sensitive operation to a spawned AI session and relies on prompt instructions to drive file-updating behavior, but it does not constrain or verify what the sub-agent is allowed to do. This creates an agentic integrity risk: untrusted daily-log content is embedded directly into the sub-agent prompt, so prompt injection in the log could influence downstream actions or cause unintended persistence behavior.
- Content
Static analysis
No suspicious patterns detected.