Back to Ruflo

Security SOTA Report — 2026-07-21

docs/dream-cycles/2026-07-21-security-sota.md

3.34.05.6 KB
Original Source

Security SOTA Report — 2026-07-21

TL;DR: Five Grade-A papers from 2026 show that authenticated agents remain hijackable via planning-phase injection and adaptive multi-turn attacks, with Neural Cryptographic Authorization as the emerging architectural solution; no major framework — LangGraph, AutoGen, CrewAI, or OpenAI Swarm — has implemented first-class prompt injection defenses despite OWASP LLM01 ranking it the #1 risk.


What's New in 2026

FindingSourceConfidence
Adaptive Adversaries: stateful multi-turn attackers break all evaluated single-round defenses across tested LLM agentsarXiv (Jul 20, 2026)A
PlanFlip: one injection into the Planner's context cascades to corrupt all downstream sub-tasks in multi-agent pipelinesarXiv (Apr/Jul 2026)A
Neural Cryptographic Authorization: authenticated agents remain hijackable via IPI even after identity verification; mandatory cryptographic gate before any agentic action is the proposed fixarXiv (Jul 16, 2026)A
MemPoison: adversarial content injected via standard channels persists across turns; compositional + trigger-conditioned attacks bypass all evaluated defensesarXiv (Jul 16, 2026)A
Agent Skill Security: supply-chain attacks operate at repository admission and semantic retrieval stages — before a skill ever executesarXiv (Jul 15, 2026)A
CrewAI 1.15.1 (Jul 2026) patched a real SSRF open-redirect bypass in its built-in scraping tool, enabling agent-driven exfiltration of internal endpoints via crafted URLsGitHub releases (Jul 2026)B

Ruflo Current Capability

AreaState
Prompt injection (OWASP LLM01)@claude-flow/security InputValidator + ADR-178 RepE detection (Proposed, not shipped)
Memory securityADR-178 VMG proposed but unimplemented; MemPoison compositional blindspots unaddressed
Planning-phase guardsNone — no injection validation at the planner boundary before sub-task distribution
Adaptive adversary resistanceNone — hooks are single-round; no stateful attacker modeling
Cryptographic action gatingNone — no Neural Cryptographic Authorization layer in hooks or security module
Skill supply-chain admissionADR-165 proposes plugin signing; skill-lifecycle admission checks absent
Code execution sandboxingNone equivalent to AutoGen Docker-default

Competitor Comparison

FrameworkVersion (Jul 2026)Prompt Injection DefenseSandboxingSupply-Chain AuthSSRF FixHuman Gate
Ruflo3.32.9RepE ADR-178 (proposed)NoneADR-165 (proposed)NoneNone
LangGraph1.2.9NoneNoneNoneNoneNone
AutoGenv0.7.5NoneDocker-default ✓NoneNoneapproval_func
CrewAI1.15.5NoneNoneSkill registry auth ✓Fixed ✓None
OpenAI SwarmNo releasesNoneNoneNoneNoneNone

OWASP LLM Top 10 (2025): LLM01 Prompt Injection, LLM03 Supply Chain, LLM06 Excessive Agency, LLM08 Vector/Embedding Weaknesses — none of the four competitors have first-class LLM01 defenses at the orchestration layer.


Benchmarks

ClaimValueGradeSource
Multi-turn adaptive attacks defeat all single-round defendersDefender success rate collapses under stateful pivot attackerAarXiv Adaptive Adversaries (Jul 20, 2026)
PlanFlip cascade amplification via Planner context injectionCorrupts all downstream sub-tasks from a single injected payloadAarXiv PlanFlip (Apr/Jul 2026)
MemPoison compositional trigger attacks bypass all defensesStructural blind spots confirmed across full retrieval/execution lifecycleAarXiv MemPoison (Jul 16, 2026)
OWASP LLM01 = #1 risk; no framework ships orchestration-layer defensesAudit of LangGraph, AutoGen, CrewAI, OpenAI Swarm — all absentAOWASP 2025 + competitor audit (Jul 2026)

SOTA Proof & Witness

Session commit: 26c35b59b40a0a95b286ccf5ac675a15edcc995f Report SHA-256 (pre-stamp): 5d839212c46a9600b6c329e95a91ffa058144c58381387564ecd3e87858b9bf6 Witness stamp: e455247f1a0e55eff837b4a8cecf6ac59aa4ec709fa6ffb28b2c6638251a0baa

Verifier: sha256sum the pre-stamp file, concat with session commit, sha256 → must equal Witness stamp.


  1. Implement Neural Cryptographic Authorization (NCA) gate (ADR-320): before pre-task hooks allow tool execution, compute a neural intent signature and require cryptographic proof of authorization chain. Pattern from arXiv Jul-16-2026 "From Neural Intent to Cryptographic Authorization: Governing Agentic Workflows". Scope: @claude-flow/security NCA middleware + @claude-flow/hooks pre-task integration. Ships alongside ADR-178 RepE detection as complementary layers (detect + prevent).

  2. Harden the planning-phase against PlanFlip injection: wrap the task planner (hooks pre-task) with an isolation boundary — strip any injected <tool_call> or <agent_instruction> patterns from retrieved context before distribution to sub-agents. Labels: OWASP LLM01 + LLM06 mitigation.

  3. Ship ADR-178 RepE IPI detection (filed 2026-07-06, still Proposed): the MemPoison paper confirms structural blindspots worsen over time. Assign to @claude-flow/hooks background worker audit — add hidden-state entropy monitoring at the tool-input position. Target: one sprint to move ADR-178 from Proposed to Accepted with a concrete implementation PR.