Skip to content

v1.0.0-core - Sellable Core (Production / GA)

Choose a tag to compare

View all tags
@mananjp mananjp released this 28 Aug 21:33
v1.0.0-core
9d2003c

v1.0.0-core — Core (Production / GA)

This release captures the production-grade core that is ready for design-partner and pilot engagements. All build, lint, typecheck, unit, integration, security, Docker, frontend, and e2e CI gates are green on this tag.

Live evidence (B2 re-run)

Reproduced on 2026-08-28T21:39:29Z against the live deployment https://a2a-firewall1.onrender.com (workspace 36349e22-c6a4-43ab-a010-f6063adfb122) via the a2a-firewall-sdk (see live case-study report):

  • Clean 2-hop pipeline (Planner → Researcher → Summarizer): ALLOWED, risk 0.0 both hops.
  • Prompt-injection / instruction-smuggling attempt: BLOCKED (risk 1.0, 6 violations: forbidden_pattern, unrestricted_access, ips_signature_prompt_injection, semantic prompt_injection).
  • SQL-injection / credential-exfiltration attempt (UNION SELECT vs internal workspace store): BLOCKED (risk 0.95).
  • Deterministic pipeline latency (in-process, Section 4 of report): p50 0.63 ms, p95 1.29 ms, p99 2.26 ms; false-positive rate 0.0% across the 211-fixture benign corpus.

SDK is published and verified on both registries at v0.2.0: PyPI a2a-firewall-sdk and npm a2a-firewall-sdk.

What's included (Core — GA)

  • 5-layer detection pipeline: preflight → agent permission / non-amplification → rules → IPS/PII/CVE → Groq semantic → decision, with identity (Ed25519) and delegation (macaroon) verification.
  • Non-amplification (confused-deputy) defense across delegation hops — a child request is blocked if it widens risk or switches task_type beyond the parent token's caveats.
  • Identity & delegation: agent cards, workspace root keys, macaroon-style attenuable tokens, delegation-chain audit log.
  • Enterprise governance: RBAC (workspace members, custom roles), SCIM provisioning, data-retention policies, spend limits, compliance packs (RBI/DPDP/HIPAA/PCI-DSS/GDPR/CCPA), SOC alerting, IP allowlist / rate limiting.
  • Network security: agent-to-agent and CIDR access rules, IPS (CVE/CVSS, agent auto-suspension).
  • MCP + proxy: Layer 2 tool-governance gateway and Layer 3 eBPF/process egress guard ships as R&D (see Roadmap below).
  • SDK (TypeScript + Python), frontend dashboard, local benchmark + case-study tooling.
  • CI is fully green: ruff lint, ruff format, mypy strict, pytest unit, integration (postgres+jaeger), e2e, security audit (pip-audit/bandit), Docker build, frontend lint/typecheck/test/build.

Note: v1.0.0-core is the Core milestone. Transparent proxy, eBPF/process egress guard, and the full-stack benchmark remain Active R&D (root/kernel tooling, not yet GA-reviewed). See README "Core vs Roadmap".

Since v0.2.0 (highlights)

  • Case study live execution — report + terminal output captured against a live Render deployment, plus an automated runner script.
  • Enterprise governance batch — RBAC, SCIM, retention, spend, compliance, SOC, IPS.
  • MCP tool-governance gateway and process egress guard (R&D).
  • Transparent TLS-intercepting proxy (R&D) with dynamic CA + OS CA trust.
  • 3-layer full-stack benchmark, zero-FP egress-guard whitelisting, platform support matrix.
  • SDK-TS renamed to @mananpanchal/a2a-firewall-sdk.
  • Quality fixes on this tag: SQLAlchemy 2.0 Mapped[...] type annotations → mypy fully clean; ruff lint + format clean; /v1/firewall/inspect now accepts delegation_token so non-amplification enforcement actually runs over HTTP; accurate case-study claims (TPR/FNR documented as R&D).
Assets 2
Loading