Skip to content

[Dream Cycle 2026-05-31] security: MCP auth gap + AIRGuard authorization propagation#2249

Draft
ruvnet wants to merge 1 commit into
mainfrom
dream/2026-05-31-security
Draft

[Dream Cycle 2026-05-31] security: MCP auth gap + AIRGuard authorization propagation#2249
ruvnet wants to merge 1 commit into
mainfrom
dream/2026-05-31-security

Conversation

@ruvnet
Copy link
Copy Markdown
Owner

@ruvnet ruvnet commented May 31, 2026

Rotation

Field Value
SLOT 1
DEEP security
SCAN intelligence, swarm
Date 2026-05-31
Session commit 05bb9cf7ed1aa30313c42553ca7c49e7574af341

Changes

  • v3/docs/adr/ADR-144-agent-authorization-propagation.md — new ADR: AgentAuthorizationPropagator + MCP auth validator
  • v3/docs/adr/README.md — added ADR-144 row
  • v3/docs/research/dream-cycle-2026-05-31-security.md — SOTA report with witness stamp

ADR

ADR-144 — Agent Authorization Propagation and MCP Authentication Enforcement (Status: Proposed)

Distinct from ADR-131 (ToolOutputGuardrail, content-layer screening):

  • ADR-131: screens WHAT agents receive as content
  • ADR-144: controls WHO agents can act as across delegation chains

Key targets:

  • v3/@claude-flow/security/src/authorization/propagator.ts
  • v3/@claude-flow/cli/src/mcp/auth-validator.ts

Research Summary

Three Grade A findings tonight:

  1. MCP auth gap (arXiv:2605.22333): 40.55% of 7,973 live MCP servers have zero auth; 96.6% of OAuth servers have ≥1 flaw. Ruflo currently has no runtime MCP server authentication validator.

  2. AIRGuard (arXiv:2605.28914): Per-action least-privilege enforcement drops attack success 36.3% → 5.5% (−85%). Ruflo has session-level auth only.

  3. Authorization propagation (arXiv:2605.05440): When agent A delegates to agent B via SendMessage, B can escalate to tools A was not authorized to call. Not solvable by RBAC/ABAC — scope must travel with the message.

Intelligence scan: Single-agent false-continue rate on infeasible tasks is 73.9% (arXiv:2605.28532, Grade A). Feasibility pre-check in route hook is a zero-ADR fix.

Swarm scan: Event-triggered consensus (arXiv:2604.06813, Grade A) reduces idle-agent token burn while maintaining ≥99% completion. Enhancement to ADR-132.


Gist / Witness

Research file: v3/docs/research/dream-cycle-2026-05-31-security.md
Issue: #2248
Report SHA-256: a7097af834cb47d04ec6c3a89b8698a90a003f82de746b82d78b6548abe24af2
Witness stamp: 3e9b27fbe7f1bc645ce09a95dd015a325d2ecfb618ca2db7f49b25a4df8d08fe


Merge Policy

Do not self-merge. Leave for human review. This PR contains an architectural proposal (ADR-144) that requires deliberate evaluation before implementation begins.

Closes companion issue #2248.


Generated by Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants