The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides Python templates and examples that execute high-privilege system commands using the subprocess module. Examples include systemctl restart for service management, find ... -delete for file system operations, and various kubectl commands for container orchestration.
[COMMAND_EXECUTION]: Chaos engineering scripts in references/incident-chaos.md utilize tc (traffic control) to inject network latency and iptables to create network partitions, requiring root-level capabilities.
[REMOTE_CODE_EXECUTION]: The AutomatedRunbook class in references/automation-toil.md performs unsafe command execution using subprocess.run(step.command, shell=True). This implementation lacks input sanitization and could allow arbitrary code execution if the runbook steps are generated from untrusted data or user input.
[EXTERNAL_DOWNLOADS]: Automation logic is driven by data fetched from external network endpoints, such as http://prometheus:9090 in SKILL.md and http://localhost:8080/health in references/automation-toil.md, which could be exploited through server-side request forgery (SSRF) or data poisoning to trigger unintended system actions.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection where malicious instructions could be embedded in the data processed by the agent to influence execution of privileged commands.
Ingestion points: Internal Prometheus API (SKILL.md), local health endpoints (references/automation-toil.md), and potentially service logs or incident reports.
Boundary markers: None identified; there are no instructions to the agent to delimit or ignore instructions embedded in the external metric data or logs.
Capability inventory: High-privilege access to kubectl, systemctl, iptables, and tc via shell execution.
Sanitization: Absence of shell argument validation or sanitization in the provided automation templates.

sre-engineer