Security, safety, and efficiency rules for AI agents calling SigmaShake MCP (Model Context Protocol) server tools. Blocks rule injection attacks, API key exposure in audit logs, accidental hub publishing, and token-wasting query patterns.
14 rules · 3 files
ssg hub pull rules-mcpAvailable on the SigmaShake Hub — the open registry for AI agent governance rules. Compatible with Claude Code, GitHub Copilot, Cursor, Windsurf, Aider, and any AI coding agent using the ssg hook protocol.
| Rule | Decision | Severity | Description |
|---|---|---|---|
no-eval-inline-allow |
DENY | error | Blocks inline rules containing ALLOW (rule injection attack) |
no-eval-inline-force |
DENY | error | Blocks inline rules containing FORCE (rule injection vector) |
no-hub-submit-bad-url |
DENY | error | Requires valid https://github.com/owner/repo URL for hub submissions |
no-token-in-eval-input |
DENY | error | Blocks GitHub tokens and API keys in eval payloads (persisted in audit log) |
ask-hub-submit |
ASK | warning | Human approval before publishing a ruleset to the hub |
| Rule | Decision | Severity | Description |
|---|---|---|---|
ask-large-audit-query |
ASK | warning | Confirms need for audit queries with limit >= 50 |
ask-eval-empty-input |
ASK | warning | Catches eval calls with empty tool name or empty input object |
force-rule-id-format |
FORCE | warning | Enforces kebab-case rule_id format for rules_get |
log-eval-call |
LOG | info | Logs every eval invocation for session observability |
log-audit-no-filter |
LOG | info | Notes unfiltered audit queries and recommends filtering |
| Rule | Decision | Severity | Description |
|---|---|---|---|
log-status-call |
LOG | info | Tracks health checks; reminds to avoid polling |
log-rules-list-call |
LOG | info | Tracks rules_list calls; recommends caching |
log-rules-get-call |
LOG | info | Tracks individual rule fetches; recommends batching |
log-audit-query |
LOG | info | Tracks audit queries; notes point-in-time nature |
MCP tools expose governance infrastructure directly to AI agents. Without guardrails:
- Rule injection: An agent could craft inline ALLOW rules passed to
sigmashake_evalto override existing DENY rules, defeating the entire governance layer - Secret leakage: GitHub tokens passed through eval inputs are persisted in the audit log as plaintext — tokens belong only in
sigmashake_hub_submit - Accidental publishing: Hub submissions without human review could publish broken or internal rulesets publicly
- Context waste: Unfiltered audit queries and repeated status checks consume LLM tokens without adding value
These rules enforce defense-in-depth: security rules block dangerous patterns outright, safety rules add human checkpoints for risky operations, and efficiency rules provide observability without interrupting the agent.
Part of the SigmaShake Hub — open-source governance rules for AI coding agents.
Install the ssg CLI to enforce these rules: npm install -g @sigmashake/ssg