Back to skill

Security audit

Lena Learning

Security checks for vulnerabilities and agentic risk

Overview

This skill is a self-learning memory tool, but it automatically persists conversation-derived content and may update agent instruction files without clear user review or limits.

Review this before installing if you do not want conversations, corrections, preferences, or project context saved across sessions. It should be changed to require explicit approval before any memory write, avoid writing conversation-derived content into AGENTS.md or TOOLS.md, exclude secrets and sensitive personal data, and provide a clear way to inspect and delete saved memory.

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)

T02 · Agent Memory Poisoning

Error
Location
workflows/self-improvement.md:6
Finding

Untrusted Conversation Content Can Poison Persistent Agent Memory

Content
View full analysis
Scanne die letzten Nachrichten nach Feedback Extrahiere Korrekturen ("Nein, das ist falsch...", "Besser...") Identifiziere neue Präferenzen Schreibe in memory/YYYY-MM-DD.md Update MEMORY.md wenn wichtig ``` `workflows/self-improvement.md:20-28`: ```text Erkenne Korrektur-Pattern Speichere in memory/YYYY-MM-DD.md mit Zeitstempel Prüfe ob Regel/Präferenz betroffen Update AGENTS.md oder TOOLS.md wenn nötig Bestätige Verbesserung Korrektur dokumentiert + zukünftige Verbesserung ``` `SKILL.md:18-26`: ```text ### 3. Feedback Loop - Korrekturen sofort speichern - recurring patterns merken - bessere prompts entwickeln ``` `SKILL.md:43-49`: ```text Aktualisiere Memory Files
- memory/YYYY-MM-DD.md: Raw notes - MEMORY.md: Langzeit-Wissen - USER.md: Präferenzen - TOOLS.md: Environment-Notes
``` `SKILL.md:58-69`: ```text Feedback-Loop
- Wenn Thomas mich korrigiert -> sofort speichern - Wenn etwas nicht funktioniert -> dokumentieren - Wenn etwas gut funktioniert -> merken
## Wann aktivieren? - Am Ende jeder Session - Nach jeder Korrektur durch Thomas ``` ### Technical Analysis The Skill directs the agent to inspect conversation messages, identify apparent corrections or preferences, and persist them in files that influence future sessions. These destinations incl ...[truncated 2985 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (15)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The objective states that the agent stores insights, corrections, and preferences for future responses, but it does not warn users that conversation-derived data will be persisted to files. This creates a transparency and consent failure, increasing the risk of retaining sensitive or personal information without the user's informed approval.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The entire skill description and operational instructions are written in German and imply the skill's behavior and interaction model are fixed to that language, with no indication that users may choose another language or that the locale restriction is intentional and justified. This can violate language or locale policy where user choice or explicit documentation is required.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill explicitly instructs the agent to learn from every conversation and store insights, corrections, and preferences across sessions. This creates a direct data retention risk because broad, free-form memory capture can include secrets, personal data, or sensitive contextual details that later become exposed or misused.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The memory system specifies daily logs, long-term memory, and user preference files, which encourages systematic collection of conversation-derived content. Because the categories are broad and unconstrained, the agent may store far more user data than necessary, increasing the chance of privacy violations and cross-session leakage.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The post-conversation routine requires updating raw notes, long-term memory, and preference files after each conversation. In context, this reinforces automatic retention as a default behavior, making accidental storage of sensitive user content more likely and increasing the blast radius of any later memory disclosure.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger section activates the skill at the end of every session, after every correction, for significant decisions, and daily via a heartbeat. Those conditions are broad and subjective, which can cause the agent to run persistence behavior unexpectedly and collect or write data outside a clearly consented workflow.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

Line L15 lists "Deutsch" as a preference, which is a natural-language locale constraint. Because the file does not indicate that the user can choose another language or opt in to German responses, this can violate the policy against forcing a specific language without user consent.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

This workflow instructs the agent to scan recent messages for feedback and write learnings, corrections, and preferences into dated memory files without any notice or consent mechanism. Persisting user-derived data across sessions creates privacy and retention risks, especially when the user may not expect that conversational content is being stored as long-term memory.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The instruction to persist feedback and preferences across sessions establishes long-term storage of user-derived behavioral data. Without consent, minimization, or sensitivity filtering, this can retain private information and also allow adversarially supplied preferences or false corrections to shape future behavior persistently.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The correction-handler uses broad natural-language triggers like recognizing a correction pattern and then immediately saving it, which can cause unintended activation from casual dialogue, quoted text, or adversarial prompts. Because the workflow persists the extracted content and may propagate it into AGENTS.md or TOOLS.md, a misfire can turn ordinary conversation into durable configuration or memory changes.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The workflow directs immediate storage of user corrections and allows those corrections to trigger updates to AGENTS.md or TOOLS.md without a user-facing warning. This is more dangerous than simple note-taking because user-supplied content can influence operational instructions or tool-related documentation, creating a path for persistent prompt injection or unsafe configuration drift.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

Immediately saving corrections with timestamps and potentially updating agent documentation turns transient user input into authoritative persistent state. In skill context, this is especially dangerous because AGENTS.md and TOOLS.md can influence future system behavior, so malicious or mistaken user content may become a durable instruction source.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The workflow continuously monitors and remembers user preference patterns, including communication style and timing, as an ongoing behavior. This creates a surveillance-like persistence model that can accumulate sensitive behavioral profiles and be manipulated by attackers who intentionally seed false preferences into conversation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The note lists 'Deutsch' as a standing preference to track, which effectively fixes the interaction language for the skill. There is no indication that the user can choose another language or that this locale restriction is optional or justified as region-specific.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
91% confidence
Finding

The template headings and placeholders use German terms such as "Neue Learnings," "Korrekturen," and "Nächste Schritte," which effectively imposes a language choice on users of the skill. The file does not indicate that German is optional or that the template is intended for a German-specific audience.

Content

No source excerpt is available for this finding.