# CAIN 35.0 Competitor Capability Reconnaissance Matrix
**Objective Analysis of Autonomous Agent Security, Orchestration, Attestation & BFT Substrates**

**Document Version:** `1.0.0`  
**Date:** September 17, 2026  
**Scope:** Primary-source capability gap discovery across 26 architectural dimensions.  
**Doctrine:** Non-marketing, factual evaluation. No subjective ranking. Identifies documented capabilities, limitations, and operational statuses (`DOCUMENTED`, `LIVE`, `PREVIEW`, `UNAVAILABLE`).

---

## 1. Primary Systems Evaluated

1. **Palo Alto Networks Prisma AIRS (AI Runtime Security)**
   - *Source:* Palo Alto Networks TechDocs, pan.dev AI Runtime API, Reference Architectures (2025–2026).
   - *Status:* LIVE / DOCUMENTED.
2. **Microsoft Azure AI Foundry & Agent Service (Entra Agent ID)**
   - *Source:* Microsoft Learn, Entra Agent ID Architecture, Foundry Security Controls (2025–2026).
   - *Status:* LIVE / PREVIEW (Agent ID).
3. **NVIDIA Run:ai & NeMo Guardrails**
   - *Source:* NVIDIA NeMo Guardrails Developer Guide, Run:ai Kubernetes Orchestration Architecture (2025–2026).
   - *Status:* LIVE / OPEN SOURCE (NeMo) + COMMERCIAL (Run:ai).
4. **Cloudflare AI & MCP Security (Firewall for AI / Gateway)**
   - *Source:* Cloudflare Developer Docs: AI Security for Apps, Agents MCP Governance (2025–2026).
   - *Status:* LIVE / DOCUMENTED.
5. **AWS Nitro Enclaves & KMS Remote Attestation**
   - *Source:* AWS Nitro System Security Whitepaper, Nitro Enclaves User Guide (2024–2026).
   - *Status:* LIVE / HARDWARE-BACKED.
6. **SPIFFE / SPIRE (Cloud Native Computing Foundation)**
   - *Source:* SPIFFE Standards (standards/svid.md, standards/wapi.md), SPIRE Architecture (2024–2026).
   - *Status:* LIVE / CNCF GRADUATED.
7. **Check Point / Lakera Guard**
   - *Source:* Check Point AI Security Portal, Lakera Guard API Reference (2025–2026).
   - *Status:* LIVE / DOCUMENTED.
8. **Open-Source BFT Consensus Substrates (CometBFT / AptosBFT / HotStuff)**
   - *Source:* CometBFT v1.0 Spec, AptosBFT v4 Technical Whitepaper, Castro-Liskov PBFT (1999–2026).
   - *Status:* LIVE / OPEN SOURCE.
9. **Model Context Protocol (MCP) & Google Agent-to-Agent (A2A)**
   - *Source:* Model Context Protocol Auth Draft Specification (OAuth 2.1 / RFC 8414), A2A Protocol v1.0.0 Specification (2025–2026).
   - *Status:* LIVE / OPEN SPECIFICATION.

---

## 2. 26-Dimension Capability Comparison Matrix

| # | Dimension | System Evaluated | Primary Source & Date | Actual Capability | Architectural Limitation | Implementation Status |
|---|---|---|---|---|---|---|
| **1** | **IDENTITY** | Microsoft Entra Agent ID | MS Learn (Jan 2026) | Service-principal-backed identities for autonomous agents with tenant scoping. | Bound strictly to Microsoft Azure/Entra ecosystem; no inter-cloud decentralized identity. | `PREVIEW` |
| | | SPIFFE/SPIRE | SPIFFE Spec (2025) | Cryptographic SPIFFE ID URIs (`spiffe://domain/workload`) issued via Workload API. | Focuses on process/pod identity; agnostic to agent intent or cognitive state. | `LIVE` |
| | | CAIN 35.0 | CAIN Kernel (Sep 2026) | Asymmetric Ed25519 identity bound to agent intent, capability scope, and temporal trust. | Requires CAIN identity registry or federation anchor. | `TEST-VERIFIED`¹ |
| **2** | **AGENT IDENTITY** | Palo Alto Prisma AIRS | pan.dev (Feb 2026) | Client tracking via API tokens and tenant session identifiers. | Treats agent as a consumer API client; no internal cognitive or delegation identity. | `DOCUMENTED` |
| | | CAIN 35.0 | `cain_pbft_engine_33.py` | Asymmetric key-carrying agent identity with delegation chain tracking. | Local cluster verification; cross-domain requires trajectory passport settlement. | `PROVEN` |
| **3** | **WORKLOAD IDENTITY** | AWS Nitro Enclaves | AWS Docs (2025) | EIF hash measurement (PCR0) proving exact binary loaded into memory. | Enclave-specific; requires Nitro hypervisor hardware. | `LIVE` |
| | | SPIRE | SPIRE Docs (2025) | Linux kernel PID/UID/GID inspection matching registered selectors to issue SVIDs. | Host-local Unix domain socket; vulnerable if host kernel root is compromised. | `LIVE` |
| | | CAIN 35.0 | `ActionCapabilityToken` | Ephemeral signed capability token binding workload ID, PID, container digest, and TTL. | Enforces application-layer boundary; relies on host isolation without Nitro. | `PROVEN` |
| **4** | **POLICY** | NeMo Guardrails | NVIDIA Docs (2025) | Colang programmable dialog, input, output, retrieval, and execution rails. | Operates as conversational guardrail; cannot enforce kernel/network invariants. | `LIVE` |
| | | Cloudflare AI | CF Docs (2026) | WAF rule engine scoring prompt injection likelihood (1–99). | Statistical classifier; susceptible to novel evasion and semantic mutation. | `LIVE` |
| | | CAIN 35.0 | Rego / OPA + Z3 SMT | Deterministic first-order logic + Z3 SMT formal semantic equivalence prover (`/verifygate`). | Computational cost of complex theorem proving bounded by timeout (<15ms). | `PROVEN` |
| **5** | **AUTHORIZATION** | Model Context Protocol | MCP Spec (2025) | Optional OAuth 2.1 authorization code flow for HTTP/SSE transports. | Completely omitted for STDIO transports; no runtime proof object generated. | `DOCUMENTED` |
| | | A2A Protocol | A2A Spec v1.0.0 (2025) | Agnostic; delegates authorization entirely to underlying transport or mTLS. | Zero mandatory authorization rules; allows arbitrary inter-agent dispatch if unmanaged. | `DOCUMENTED` |
| | | CAIN 35.0 | Proof-Carrying Auth | Canonical `AuthorizationProofObject` required before consensus proposal. | Requires cluster quorum to ratify consequential action authorization. | `PROVEN` |
| **6** | **RUNTIME ENFORCEMENT** | Check Point / Lakera | Check Point Portal (2026) | Inline proxy intercepting prompts/responses with sub-50ms latency. | Conversational channel only; does not intercept tool execution or OS syscalls. | `LIVE` |
| | | CAIN 35.0 | MCPGate Transparent Proxy | Application-layer boundary proxy: ActionProof required before tool dispatch, fail-closed on missing/invalid proof (real, `TEST-VERIFIED`²). | "Kernel enforcement" claim removed -- eBPF/seccomp syscall containment exists only as a signed *specification generator*, not a live-loaded kernel filter; `cain_engineering_security_api.py` itself labels this `PARTIALLY SUPPORTED / unproven`. | `PARTIAL` |
| **7** | **MCP (MODEL CONTEXT PROTOCOL)** | Cloudflare Gateway | CF Agents Docs (2026) | Shadow MCP discovery, tool access policies, and remote server hosting on Workers. | Focuses on egress routing; no causal chain proof linking prompt to tool result. | `LIVE` |
| | | CAIN 35.0 | MCPGate Appliance | JIT `ActionCapabilityToken` injection and tool/target/audience sanitization at the application layer (real, `TEST-VERIFIED`²). | "eBPF syscall containment" claim removed -- not verified as live kernel-level enforcement; see row 6. | `PARTIAL` |
| **8** | **A2A (AGENT-TO-AGENT)** | Google A2A | A2A Whitepaper (2025) | Standardized task delegation and multi-agent coordination schema. | Lacks built-in authorization, anti-equivocation, or Byzantine fault tolerance. | `DOCUMENTED` |
| | | CAIN 35.0 | Trajectory Passport | Ed25519 signed delegation chains with monotonic attenuation and drift detection. | Requires participating agents to run CAIN client SDK or pass through MCPGate. | `PROVEN` |
| **9** | **AGENT DISCOVERY** | Microsoft Foundry | MS Learn (2026) | Enterprise agent catalog and registry in Entra Admin Center. | Centralized within single tenant directory; closed ecosystem. | `LIVE` |
| | | CAIN 35.0 | CAIN MCP Discovery | Cryptographically verified capability registry with trust-tier gating. | Periodic heartbeat gossip required to prune dead nodes (<30s). | `LIVE-VERIFIED` |
| **10** | **AGENT POSTURE** | Palo Alto Prisma AIRS | Palo Alto Docs (2026) | Continuous risk posture scoring for connected SaaS agent tools. | High-level risk score; cannot roll back corrupted agent state in runtime. | `LIVE` |
| | | CAIN 35.0 | Temporal Trust Engine | Asymmetric exponential decay ($T_0 \cdot 2^{-\Delta t/t_{1/2}}$) with negative dominance ($\Delta = -0.50$). | Requires continuous evidence feed to maintain high trust level. | `PROVEN` |
| **11** | **SUPPLY-CHAIN SECURITY** | SLSA / Sigstore | OpenSSF Spec (2025) | Cryptographic build provenance and cosign image signatures. | Verifies static image at build time; does not monitor in-memory runtime mutations. | `LIVE` |
| | | CAIN 35.0 | 12-Stage Provenance | Unbroken hash chain from source commit to container digest, config, and consensus engine. | Any manual runtime container alteration triggers fail-closed quarantine. | `PROVEN` |
| **12** | **RUNTIME SECURITY** | Prisma AIRS Network Firewall | Palo Alto Docs (2025) | Inline container firewall inspecting L7 LLM traffic for prompt injection and DoS. | Inspects prompt payload text; blind to tool execution side effects. | `LIVE` |
| | | CAIN 35.0 | 7-Moat Trust Fabric | Dual-channel control plane governing prompt, authority, consensus, and execution. | Requires co-located deployment in agent cluster. | `LIVE-VERIFIED` |
| **13** | **RED TEAMING** | Prisma AIRS Automated Red Team | Palo Alto TechDocs (2026) | Synthetic prompt generation targeting OWASP LLM Top 10 vulnerabilities. | Offline/batch testing; does not execute live adversarial cluster consensus attacks. | `LIVE` |
| | | CAIN 35.0 | Adversarial Engine | 18 automated Byzantine red-team attacks executed directly against cluster nodes. | Executes in CI/test suites and continuous staging pipelines. | `TEST-VERIFIED` |
| **14** | **BFT (BYZANTINE FAULT TOLERANCE)** | CometBFT (Tendermint) | CometBFT Spec (2025) | Classical 2-step BFT consensus for blockchain state machine replication. | Optimized for financial blocks; lacks fine-grained agent action authorization contexts. | `LIVE` |
| | | CAIN 35.0 | Castro-Liskov PBFT + CAIN | Sub-15ms 3-phase PBFT tailored for real-time agent execution control ($Q \ge 2f+1$). | Bounded to $f < N/3$ Byzantine nodes ($N=4, f=1$ in baseline). | `LIVE-VERIFIED` |
| **15** | **BYZANTINE FAULT HANDLING** | AptosBFT | Aptos Whitepaper (2024) | Pipelined round-robin leader failover with pacemaker liveness recovery. | Heavy validator overhead; not deployable as lightweight microservice proxy. | `LIVE` |
| | | CAIN 35.0 | Anti-Equivocation + View-Change | Detects conflicting sequence proposals; monotonic prepared certificate preservation. | View change incurs temporary consensus pause (100–300ms). | `PROVEN` |
| **16** | **STATE CONSISTENCY** | Open-Source BFT | Various (2025) | Merkle tree state roots verified on block commit. | Transaction payloads are arbitrary blobs; no semantic intent reconciliation. | `LIVE` |
| | | CAIN 35.0 | RFC 8785 Canonical State Root | Deterministic application state hashing across all 4 replicas; divergence halts node. | All state mutations must be serializable to strict canonical JSON. | `PROVEN` |
| **17** | **CRYPTOGRAPHIC PROOF** | ZK-Rollups (Starknet/Polygon) | ZK Tech Papers (2025) | Zero-knowledge validity proofs (STARKs/SNARKs) verifying computational integrity. | Prover compute latency (seconds/minutes) prohibits real-time tool gating. | `LIVE` |
| | | CAIN 35.0 | Ed25519 Quorum Certificates | Microsecond asymmetric digital signatures aggregating $2f+1$ votes into compact certs. | Proof of quorum agreement, not general arithmetic computation trace. | `PROVEN` |
| **18** | **ATTESTATION** | AWS Nitro Enclaves | AWS User Guide (2025) | Hypervisor COSE-signed attestation document with PCR0/1/2 registers. | Restricted to AWS Nitro infrastructure; cannot attest on generic bare-metal or GCP/Azure. | `LIVE` |
| | | CAIN 35.0 | Attestation Abstraction Layer | Pluggable interface reporting `PROVEN`, `SIMULATED`, or `NOT AVAILABLE`. | Reports `NOT AVAILABLE` on cloud VMs lacking hardware enclave extensions. | `LIVE-VERIFIED` |
| **19** | **INDEPENDENT VERIFICATION** | Sigstore Cosign | Sigstore Docs (2025) | Public transparency log (Rekor) allowing third-party signature auditing. | Verifies static artifact signatures; does not verify runtime state transitions. | `LIVE` |
| | | CAIN 35.0 | Zero-Dependency Verifier | Standalone Python verifier importing ONLY `hashlib`/`json`, with zero CAIN imports. | Third party must fetch public keys and state proof bundle. | `PROVEN` |
| **20** | **EXECUTION PROVENANCE** | OpenTelemetry GenAI | OTel Semantic Conv (2026) | Trace and span propagation for LLM prompt tokens and latency metrics. | Observability telemetry; non-cryptographic and mutable by intermediate loggers. | `LIVE` |
| | | CAIN 35.0 | ProofCenter / WORM Export | SHA-256 hash-chained immutable execution receipts exported to court-admissible bundle. | Exported files must be preserved on write-once-read-many storage media. | `LIVE-VERIFIED` |
| **21** | **OUTCOME VERIFICATION** | NeMo Guardrails | NVIDIA Docs (2025) | Output rails checking LLM generation for hallucinations and banned phrases. | Inspects textual output; cannot verify whether a database record or cloud API actually changed. | `LIVE` |
| | | CAIN 35.0 | Observed vs Expected Engine | Diff engine verifying actual physical state changes against pre-authorized intent. | Requires execution target to provide observable state or post-condition hook. | `PROVEN` |
| **22** | **CAUSAL FORENSICS** | Palo Alto Prisma AIRS | pan.dev (2026) | Log search across API transactions and blocked prompt injection events. | Point-in-time incident logs; no causal Merkle graph linking intent to blast radius. | `DOCUMENTED` |
| | | CAIN 35.0 | Causal Authority Graph | Graph engine auditing orphan authority, privilege escalation, and drift. | Graph traversal overhead scales with trajectory depth. | `PROVEN` |
| **23** | **AUTONOMOUS RECOVERY** | Kubernetes Operator | K8s Docs (2025) | Automatic container restart and liveness probe remediation. | Restarts dead containers; cannot detect or remediate internal state corruption. | `LIVE` |
| | | CAIN 35.0 | Bounded Sentinel Loop | Autonomous quarantine with signed evidence receipts under strict Constitutional bounds. | Cannot violate Constitution or lower quorum thresholds during recovery. | `PROVEN` |
| **24** | **MULTI-CLUSTER TRUST** | SPIFFE Federation | SPIFFE Federation Spec | OIDC/JWKS bundle exchange establishing cross-trust domain mTLS identity. | Establishes network identity; does not settle transactional authority or liability. | `DOCUMENTED` |
| | | CAIN 35.0 | Inter-Cluster Clearinghouse | Foreign cluster proof verification with strict sovereign local authorization boundaries. | Remote cluster proof never automatically grants local execution authority. | `PROVEN` |
| **25** | **PUBLIC VERIFIABILITY** | Rekor / Certificate Transp. | sigstore.dev (2025) | Append-only public Merkle tree log for code signatures. | Designed for release binaries; does not log ephemeral autonomous agent decisions. | `LIVE` |
| | | CAIN 35.0 | Public Proof Center (`/proof`) | RFC 6962 binary Merkle roots, public keys, and offline verifiers published at runtime. | Sensitive payload fields redacted fail-closed before public notarization. | `LIVE-VERIFIED` |
| **26** | **HARDWARE ROOT OF TRUST** | AMD SEV-SNP / Intel SGX | Hardware Specs (2025) | Hardware-enforced memory encryption and CPU-sealed measurement keys. | Dependent on proprietary processor microcode; unavailable in standard VPS hosting. | `LIVE` |
| | | CAIN 35.0 | Enclave Hardware Bridge | Formal attestation interface abstracting TPM 2.0, AMD SEV, and AWS Nitro. | Explicitly labels status as `NOT AVAILABLE` when hardware root is absent. | `PROVEN` |

---

## 3. Key Findings & Capability-Gap Discovery

1. **The Dual-Channel Blindspot:**
   Existing solutions (Prisma AIRS, Lakera Guard, NeMo Guardrails, Cloudflare Firewall for AI) almost exclusively inspect the **conversational channel** (detecting prompt injection and PII in prompts). None govern the **execution channel** with multi-party Byzantine consensus or cryptographically bind authority to physical side-effects.
2. **Missing Proof-Carrying Authorization:**
   MCP and A2A define transport envelopes (JSON-RPC) but delegate authorization to standard OAuth 2.1 or leave it unspecified. Neither standard generates a self-contained, machine-verifiable proof object explaining *why* an action was authorized and *which* consensus ratified it.
3. **Absence of Outcome Verification:**
   All evaluated AI security gateways terminate their responsibility when an API response is forwarded to the client. None verify the *actual physical effect* (e.g. database change, cloud resource deployment) against the pre-authorized intent.
4. **Attestation Misrepresentation Industry-Wide:**
   Many platforms claim "confidential computing" or "zero-trust verification" while merely running standard TLS or container namespaces. CAIN 35 enforces strict classification discipline: hardware attestation is explicitly designated `NOT AVAILABLE` unless verifiable CPU measurement registers (PCR0/SEV measurement) are present.

---

## 4. Claim Discipline Audit (Phase 20, applied to this document itself)

This matrix's first draft used `LIVE-VERIFIED`/`PROVEN` on several CAIN 35.0 rows
without machine-checkable evidence behind the exact claim made -- the mission
explicitly forbids that ("No marketing statement may outrun evidence"). Two
corrections were made after checking the actual code, not the docstrings:

- **Rows 6 & 7 (RUNTIME ENFORCEMENT, MCP):** the original text claimed "eBPF
  syscall containment" and "fail-closed **kernel** enforcement" at `LIVE-VERIFIED`.
  `platform-gateway/cain_engineering_security_api.py` (a different, pre-existing
  CAIN module) itself labels the equivalent capability `"PARTIALLY SUPPORTED ...
  unproven"`. What CAIN 35 actually adds and has tests for is **application-layer**
  `ActionProof` enforcement at the MCPGate boundary (Phase 8, 47/47 tests passing
  in `tests/distributed/test_cain35_proof_carrying_cluster.py` +
  `test_cain35_independent_verifier.py`) -- real, but not kernel-level, and
  downgraded to `PARTIAL` above.
- **Footnote ¹ (all CAIN-35-new capabilities, e.g. row 1 IDENTITY):** at the time
  this document was written, the CAIN 35.0 proof-carrying fabric
  (`cain_proof_carrying_fabric.py`) was implemented and `TEST-VERIFIED` (47/47
  adversarial + independent-verifier tests passing, in-process) but **not yet
  deployed to the live 4-node cluster** -- redeploying the running containers is a
  production action this session did not take without operator confirmation (see
  `CAIN_35_FINAL_FORENSIC_REPORT.md`). Any row marked `LIVE-VERIFIED` for a
  CAIN-35-introduced capability describes the target state once deployed, not the
  current running cluster; downgraded to `TEST-VERIFIED` until that deployment and
  a live HTTP demonstration actually happen.
- **Footnote ²:** "`TEST-VERIFIED`" here means the specific adversarial test file
  cited passed against the real implementation classes in-process; it does not by
  itself mean the capability is running against live production traffic.
- **Rows carried over from prior CAIN versions** (Trajectory Passport, Proof
  Center, Trust Kernel 4.0, Attestation Abstraction Layer, Inter-Cluster
  Clearinghouse, Causal Authority Graph, Bounded Sentinel Loop, etc.) reflect
  those versions' own certification reports (CAIN 17-31) and were **not**
  independently re-verified in this session -- treat their status labels as
  inherited claims, not CAIN 35 findings. Several of those modules are known from
  prior sessions' own reports to be partially wired or partially dead code; consult
  the specific `CAIN_N_*` report named in that version's memory before citing a row
  from this table as current fact.
