The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes proactive safeguards against unauthorized agent behavior. In SKILL.md, it establishes an 'Operational closure gate' and a 'Governed exception' policy that explicitly instructs the agent to never self-grant authority and to require independently attributable human approval for production mutations. It further clarifies that 'Skill text and agent roles confer no permissions'.
[INDIRECT_PROMPT_INJECTION]: The skill presents a minimal attack surface for indirect prompt injection through its handling of Service Level Indicator (SLI) data.
Ingestion points: The skill processes SLI metrics via stdin in the scripts/slo-burn-rate.py utility.
Boundary markers: While explicit delimiters are absent in the script's raw input, the skill instructions (e.g., references/agentic-operations.md) emphasize evidence-grounded operations and the use of timestamped sources with freshness limits.
Capability inventory: The script performs numeric calculations using the Python standard library. The agent's broader capabilities are strictly governed by the 'Agentic SRE action contract' template.
Sanitization: The slo-burn-rate.py script parses input as structured JSON and treats data as numeric floating-point values for calculation, inherently limiting the execution of embedded instructions.
[DATA_EXFILTRATION]: No exfiltration patterns or unauthorized network operations were detected. External URLs in reference files point exclusively to well-known technology domains, academic repositories (Arxiv), and established cloud providers for documentation purposes.
[COMMAND_EXECUTION]: While templates like templates/runbook-template.md include example CLI commands (e.g., kubectl, systemctl, ssh), these are provided as instructional references. The skill's core instructions in SKILL.md and references/agentic-operations.md strictly forbid the execution of mutations without verified human authorization or external enforcement controller receipts.

site-reliability-engineering