Back to Ruflo

Security SOTA Report — 2026-05-31

v3/docs/research/dream-cycle-2026-05-31-security.md

3.34.04.9 KB
Original Source

Security SOTA Report — 2026-05-31

TL;DR: Three Grade A papers published in the last 10 days define a new security layer Ruflo lacks — authorization propagation across agent delegation chains — distinct from the content-screening gap (ADR-131) filed five nights ago.


What's New in 2026

FindingSourceConfidence
40.55% of 7,973 live MCP servers expose tools with zero authentication; 96.6% of OAuth-enabled servers contain ≥1 flawarXiv:2605.22333A
Runtime authority control (AIRGuard) reduces agent attack success 36.3% → 5.5% via least-privilege action-layer enforcementarXiv:2605.28914A
Dual-graph provenance defense (execution graph vs authorization intent) reduces indirect prompt injection 40% → 1%arXiv:2605.26497A
Authorization propagation across multi-agent delegation chains is not reducible to prompt injection; requires 7 structural requirements not covered by RBAC/ABACarXiv:2605.05440A
Single agents false-continue on infeasible tool tasks 73.9% of the time (intelligence scan)arXiv:2605.28532A
Event-triggered swarm consensus reduces network overhead while maintaining ≥99% task completion under agent failures (swarm scan)arXiv:2604.06813A

Ruflo Current Capability

ControlStatusGap
Content-boundary injection screeningADR-131 (Proposed)Covers WHAT agents receive — not WHO they act as
MCP tool input validationSafeExecutor + InputValidatorNo auth verification on MCP server identity
Per-action privilege enforcementNot implementedNo action → scope → allow/deny path
Delegation chain tracking in SendMessageNot implementedMessages carry no authorization scope
Execution provenance graphNot implementedNo WHAT did vs WHAT was authorized audit trail
Feasibility pre-check before tool dispatchNot implemented73.9% false-continue risk on single-agent paths

Competitor Comparison

FrameworkMCP Auth CheckingPer-Action PrivilegeAuthorization PropagationProvenance Graph
Ruflo v3.6Not implementedNot implementedNot implementedNot implemented
OpenAI Agents SDKTool availability pre-check (March 2026)Input + output + invocation guardrailsOAuth 2.0 token forwardingOTEL spans built-in
LangGraph v0.4Via LangSmith observabilityConditional edges + HITL checkpointsPartial (checkpoint-scoped)LangSmith full graph
CrewAI EnterpriseSOC 2 / HIPAA complianceRole-scoped tool permissionsPartial (role inheritance)Observability hooks
AutoGen 1.0 GASecurity patches; no MCP nativeGroupChat-level onlyNot publishedAzure Monitor integration

Benchmarks

MetricValueSourceGrade
Live MCP servers with zero auth (n=7,973)40.55%arXiv:2605.22333A
OAuth-enabled MCP servers with ≥1 flaw96.6%arXiv:2605.22333A
AIRGuard attack success reduction36.3% → 5.5% (−85%)arXiv:2605.28914A
Dual-graph injection success reduction40% → 1% (−97.5%)arXiv:2605.26497A
Single-agent infeasible-task false-continue rate73.9%arXiv:2605.28532A
Event-triggered consensus task completion≥99% under single failurearXiv:2604.06813A

SOTA Proof & Witness

FieldValue
Session commit05bb9cf7ed1aa30313c42553ca7c49e7574af341
Report SHA-256a7097af834cb47d04ec6c3a89b8698a90a003f82de746b82d78b6548abe24af2
Witness stamp3e9b27fbe7f1bc645ce09a95dd015a325d2ecfb618ca2db7f49b25a4df8d08fe
Verifiersha256sum dream-gist-2026-05-31.md (pre-witness fill) → concat session commit 05bb9cf7ed1aa30313c42553ca7c49e7574af341sha256sum → must equal witness stamp

  1. Implement AgentAuthorizationPropagator in v3/@claude-flow/security/src/authorization/propagator.ts — attach scope field to SendMessage envelope, validate each MCP tool call against the current delegation scope. ADR-144 (filed tonight) tracks this as an architectural decision.

  2. Add MCP server authentication validator in v3/@claude-flow/cli/src/mcp/auth-validator.ts — before any tool response enters agent reasoning, verify the server presented valid credentials. Even a simple allowlist check eliminates the 40.55% unauthenticated-server risk for Ruflo-managed MCP registrations.

  3. Add feasibility pre-check to the route hook — before Tier-3 dispatch, verify all required MCP tools are registered and callable. Eliminates the 73.9% false-continue rate at near-zero cost (simple registry lookup). Implementation-level — no ADR needed.