session-share

Pass

Audited by Gen Agent Trust Hub on Oct 9, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes shell scripts and external tools (agent-deck, jq, sed) to process session data. While it includes a check for session ID format in import.sh, the use of shell variables like $SESSION_ID and $IMPORT_TITLE in command lines presents a risk if input validation is bypassed.
  • [INDIRECT_PROMPT_INJECTION]: The import.sh script ingests untrusted session files created by other developers. These files contain conversation history that is directly written into the local Claude session storage (~/.claude/projects/). If a session file contains malicious instructions in the messages array, an agent resuming that session would execute them.
  • Ingestion points: import.sh reads from a user-provided JSON file path.
  • Boundary markers: None present. The content is directly converted to JSONL and placed in the session directory.
  • Capability inventory: The skill uses agent-deck to start and resume sessions, which has full access to the user's project and tools.
  • Sanitization: None performed during import. The sanitize_jsonl function is only used during export to protect the exporter's data, not the importer's safety.
  • [DATA_EXFILTRATION]: While the skill claims to sanitize data, export.sh includes a --no-sanitize flag that allows exporting raw session data including credentials. Furthermore, the sanitization relies on regex patterns (sed) which may not capture all sensitive formats or obfuscated secrets.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 9, 2026, 05:28 AM