v0.4.0GitHub

Reference

Benchmarks

Lab-tested agent-boundary benchmarks — reproducible methodology and raw JSON artifacts.

View as Markdown

Blekline publishes lab-tested benchmarks on dimensions CISO and platform teams evaluate during NHIM pilots. Every system receives identical canonical payloads. Results include raw JSON with git SHA and version pins.

Latest artifact: latest.json
Last run: 2026-07-29 · git b3584f3+ · contracts 0.4.1 · nhim-audit 0.1.1 · Lakera + Kong lab-tested

Scoring rubric

ScoreDefinition
PassBlocked or masked before model/tool execution
PartialDetected or route-level only — not structural MCP enforce
FailPayload reached execution context unchanged
N/AOutside product scope (documented reason)
SkippedAdapter not configured (missing API key or local service)

Scenario matrix (B1–B8)

IDScenarioBleklineBaselineLakeraKongOneCLI
B1Secrets in promptPassFailPass†Partial‡Partial§
B2Destructive tools/callPassFailN/APartial‡N/A
B3Session lineagePassN/AN/AN/AN/A
B4Enforce latency p990.08 msN/A306 ms†63 ms‡N/A
B5Credential in tool argsPassFailN/AN/APartial§
B6Pod egress bypassPassFailN/AN/APartial§
B7Time to first governPassN/A¶PartialPartialPartial
B8Audit artifact (0–3)Pass (3)Fail (0)Partial (2)Partial (2)Partial (2)

† Lakera Guard API — content classify; flags prompts, does not tokenize for model context.
‡ Kong serverless gateway — lab-tested; route returns 404 without upstream model route (gateway hop only).
§ OneCLI waitlist — doc-verified partial scope until local proxy available.
¶ Baseline N/A — no enforce layer; first enforced interaction never occurs (zero setup ≠ governance).

What we measure

IDICP question
B1Can raw secrets reach the model context?
B2Is tools/call with destructive args blocked at enforce hop?
B3After injection, are destructive tools blocked on contaminated session?
B4p50/p95/p99 ms on canonical enforce path
B5Are credentials masked or blocked before tool execution?
B6Can agent pods reach external HTTP without mandatory hop?
B7Minutes from zero to first enforced interaction
B8Structured allow/mask/block metadata for SIEM (score 0–3)

Reproduce locally

pnpm build:packages
cp benchmark.env.example env.benchmark   # add keys locally — never commit
pnpm benchmark:run          # full matrix
pnpm benchmark:run --quick  # Blekline + baseline only (CI)
pnpm benchmark:publish      # copy → webapp/public/marketing/benchmarks/

Environment variables

VariableUsed for
BLEKLINE_WORKSPACE_TOKENCloud mask API path (B1)
BLEKLINE_API_URLDefault https://app.blekline.com
LAKERA_API_KEYLakera Guard API
KONG_API_KEYKong Konnect / AI Gateway
KONG_AI_GATEWAY_URLLive Kong proxy URL
ONECLI_ENDPOINTOneCLI HTTP proxy (default http://localhost:8080)

Canonical payloads live in scripts/benchmarks/fixtures/payloads.json in the blekline-oss repository after sync.

Limitations

  • Lakera / Check Point Guardrails: content classification — not MCP tools/call structural enforce (B2/B3 = N/A).
  • Kong AI Gateway: route-level ACLs — stateless per request; no session lineage (B3 = N/A).
  • OneCLI: HTTP credential gateway — strong on egress keys, partial on prompt mask (B1/B3).
  • Prompt Security, Portkey, NeuralTrust: included in marketing stack comparison as doc-verified until trial APIs configured.
  • Okta, CASB: complement layers — not scored in performance benchmarks.

Webflow embeds

Marketing summary embeds (matrix, scenario cards, latency chart) ship from app.blekline.com/marketing/benchmarks/ for paste into blekline.com/benchmarks.


Next steps: Latency SLO · NHIM verification · Open workspace · Report issue