Skip to content

OWASP_LLM_TOP10_SENTINEL_MAP.md

Latest commit

 

History

History
36 lines (27 loc) · 2.02 KB

OWASP_LLM_TOP10_SENTINEL_MAP.md

File metadata and controls

36 lines (27 loc) · 2.02 KB

OWASP LLM Top 10 Sentinel Mapping

This document tracks Sentinel control coverage against OWASP LLM Top 10 (LLM01 to LLM10).

Status semantics:

  • covered: mapped controls enabled in enforce-capable mode.
  • partially_covered: mapped controls enabled but monitor/warn-only.
  • missing: mapped controls disabled.

Mapping Matrix

Risk Category Sentinel Controls Current Gap Notes
LLM01 Prompt Injection injection_scanner, prompt_rebuff Promote monitor-only deployments to block mode after evidence review.
LLM02 Insecure Output Handling pii.egress, sandbox_experimental Sandbox remains optional; enable for high-risk tool execution paths.
LLM03 Training Data Poisoning synthetic_poisoning, auto_immune Synthetic poisoning is strict-gated and should remain opt-in.
LLM04 Model Denial of Service proxy.max_body_bytes, rate_limiter Tune per-route limits for heavy internal batch paths.
LLM05 Supply Chain Vulnerabilities mcp_poisoning, provenance Expand signed policy-bundle rollout in multi-team deployments.
LLM06 Sensitive Information Disclosure pii ingress, pii_vault Vault active mode can be phased for selected flows first.
LLM07 Insecure Plugin Design canary_tools, sandbox_experimental Keep plugin hooks monitor-first before block promotion.
LLM08 Excessive Agency loop_breaker, agentic_threat_shield Strengthen identity token verification in enterprise profiles.
LLM09 Overreliance intent_drift, cognitive_rollback Auto rollback should remain controlled with explicit approvals.
LLM10 Model Theft provenance_signing, swarm_protocol Swarm is optional; provenance should be default in regulated environments.

Report Generation

node ./cli/sentinel.js compliance owasp-llm --report html --out ./owasp-llm-report.html

JSON output is also supported:

node ./cli/sentinel.js compliance owasp-llm --report json --out ./owasp-llm-report.json