other
- Location
SKILL.md:75- Finding
Unconditional Persistent Storage of Search Queries and Results
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 75-89
Vulnerability Type: Excessive Data Retention
Risk Level: MediumVulnerable Code
markdown ## Search Results Auto-cache All search results automatically saved to `memory/research/` folder: **Filename Convention:**memory/research/search-YYYY-MM-DD-HH-MM-[keyword].md
text **Saved Content:** - Search query - Timestamp - Search method used (Brave/DuckDuckGo/claude.ai) - Search results (links + summaries) - Extracted insightsTechnical Analysis
The skill requires all searches to be automatically persisted in the
memory/research/directory. Stored data includes the original query, timestamps, external results, summaries, and extracted insights. Search functionality does not inherently require persistent storage, making this retention broader than necessary for the skill's primary purpose.The documented behavior provides no consent mechanism, sensitivity screening, secret redaction, access-control requirements, retention period, or deletion procedure. Queries may contain confidential business topics, personal information, credentials accidentally supplied by a user, or other sensitive material. Search results may also contain untrusted external content, which is copied into a persistent agent-accessible location.
This is not confirmed agent memory poisoning because the documentation does not instruct the agent to treat cached content as future rules or executable instructions. The issue is instead uncontrolled retention and potential disclosure of sensitive data.
Attack Path
- A user submits a sensitive search query containing confidential terms, personal information, or an inadvertently included secret.
- The skill transmits the query to Brave, DuckDuckGo, or claude.ai as part of its declared search operation.
- Regardless of whether persistence is necessary, the skill automatically writes the query, time ...[truncated 951 chars]
- Remediation
View remediation
Remediation Suggestions
- Make caching opt-in rather than automatic, and clearly request user approval before writing search data to disk.
- Default to returning results only in the current session when persistence is not explicitly required.
- Detect and redact credentials, tokens, personal information, and other sensitive values before storage.
- Allow users to select which fields are retained; avoid storing the complete query when a sanitized identifier is sufficient.
- Store cached research outside directories used as agent memory or instruction sources to reduce the chance that untrusted web content influences future sessions.
- Apply restrictive filesystem permissions so cached data is accessible only to the intended account.
- Define a short retention period and implement automatic expiration and secure deletion.
- Provide commands to inspect and delete cached records.
- Record the cache operation and destination in the response so users know that persistence occurred.
- Treat cached web content as untrusted data and ensure it is never interpreted as agent instructions or executable code.