The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Suspected typosquatting attack identified. The installation instructions specify the package graphifyy, while the external resource link points to a repository named graphify. This discrepancy suggests an attempt to trick users into installing a potentially malicious package with a similar name.
[COMMAND_EXECUTION]: The skill establishes a persistence mechanism by appending a recurring task to the user's crontab to run a daily synchronization script. This ensures the skill's background processes continue to run indefinitely.
[CREDENTIALS_UNSAFE]: Users are instructed to hardcode and export sensitive API credentials, such as ANTHROPIC_API_KEY, directly into persistent shell configuration files like .bashrc or .zshrc.
[COMMAND_EXECUTION]: The skill dynamically creates and executes localized shell and Python scripts (sync_claude_obsidian.sh, claude_to_obsidian.py) and uses chmod +x to grant them execution privileges on the host system.
[DATA_EXFILTRATION]: The skill automates the extraction of all Claude conversation history using claude-extract --all, centralizing potentially sensitive data into a local vault where it is processed by unverified scripts.
[PROMPT_INJECTION]: The skill implements an indirect prompt injection surface by directing the agent to ingest external chat logs and architecture notes during the /resume workflow.
Ingestion points: Vault logs and architecture files defined in SKILL.md and CLAUDE.md.
Boundary markers: No boundary markers or instruction-delimiting warnings are present in the processing logic.
Capability inventory: Includes file writing, script execution, and Git operations (commit/push).
Sanitization: No sanitization or content filtering is applied to the imported chat data before it is re-ingested by the agent.

claude-code-memory-obsidian-graphify