Back to Ruflo

Security SOTA Report — 2026-06-21

docs/dream-cycle/2026-06-21-security-sota.md

3.34.06.9 KB
Original Source

Security SOTA Report — 2026-06-21

TL;DR: Six high-severity arXiv papers published in June 2026 demonstrate that Ruflo's ToolOutputGuardrail covers ASI01 detection but is missing execution-phase sandbox enforcement, cross-interaction memory fragmentation defenses, and task-scoped transactional rollback — three gaps with grade-A benchmark evidence of ≥33% attacker success rates.


What's New in 2026

FindingSourceConfidence
SafeClawBench: 291/347 sandbox harms pass semantic checks (9–44.2% per-model semantic failure rate)arXiv:2606.18356 (open benchmark, HuggingFace dataset)A
Aura Mobile: semantic firewall + privilege isolation cuts ASR from ~40% → 4.4%, raises TSR 75% → 94.3%arXiv:2602.10915 (MobileSafetyBench eval)A
FinVault: SOTA models still reach 50% ASR in financial agent scenarios despite existing defensesarXiv:2601.07853 (31 regulatory cases, 963 test cases)A
FragFuse (arXiv:2606.15609): 86.3% bypass success via long-term memory fragmentation across interactionsarXiv:2606.15609B (independent dataset, partial tool diversity)
CmdNeedle: 69–98.6% of real-world command denylists bypassable via overlooked equivalentsarXiv:2606.15549B (corpus of real denylists; cross-agent validation limited)
Skill composition: 33.6% attack success when individually-harmless skills are chained vs ~0% in isolationarXiv:2606.15242B (lab-constructed composition set)
Cordon (arXiv:2606.17573): task-scoped transactional rollback exposes cross-step violations missed by per-call guardrailsarXiv:2606.17573B (prototype; production latency not reported)
OWASP released separate Top 10 for Agentic Applications (ASI 2026): ASI01–ASI10 distinct from LLM01–LLM10OWASP ASI 2026B (OWASP official, no benchmark)
MCP-TDP: GPT-4o shows ~100% ASR in 6 high-risk tool-description-poisoning scenarios; prompt guardrails ineffectivearXiv:2605.24069B (32 cases, limited model diversity)
AutoGen 1.0 GA (Feb 2026): event-driven architecture, sandboxed tool-use; OpenAI Agents SDK (March 2026) native sandboxing + sub-agents + MCPVendor changelogsC

Ruflo Current Capability

CapabilityModuleCoverageGap
Input validation (Zod-based)InputValidatorHTTP/CLI boundariesNone — adequate
Command injection preventionSafeExecutorAllowlist-gated executionAllowlist may miss composition patterns (CmdNeedle)
Path traversal preventionPathValidatorFile-system boundariesNone — adequate
Prompt injection detection (per-call)ToolOutputGuardrailMCP/tool output, memory readsDetection-only; no execution-phase sandbox, no rollback
OWASP ASI01 (agent goal hijack)ToolOutputGuardrailPattern-match + policyCovers detection but not composition-phase attacks
OWASP ASI06 (memory poisoning)PartialNo cross-interaction trackingGap: FragFuse-style fragmentation undetected
Plugin integrityintegrity-verifierHash verificationCovers supply chain; not runtime composition
Authorization propagationauthorization/propagatorClaims-basedNo certificate-bound runtime enforcement (Sovereign Brokers pattern)
Transactional rollbackMissingCritical gap: Cordon pattern absent
Sandbox execution enforcementMissingCritical gap: ASR 291/347 per SafeClawBench

Competitor Comparison

FrameworkSandbox EnforcementMemory Poisoning DefenseTransactional RollbackOWASP ASI CoverageAudit
Ruflo v3.6Detection-only (ToolOutputGuardrail)None (cross-interaction)NoneASI01 partialIntegrity verifier
AutoGen 1.0 (GA Feb 2026)Native sandboxed tool-use (event-driven)Not documentedPartial (event replay)ASI01–ASI03 partialEnterprise observability
OpenAI Agents SDK (March 2026)Native sandboxing, Codex-style filesystemNot documentedNone documentedASI01 (guardrail API)Tool-call logs
LangGraph 0.3.xNo native sandbox; PostgresSaver checkpointingNot documentedVia checkpointer (partial)ASI01 via customState snapshot
CrewAI 0.105No native sandboxNot documentedNoneASI01 partialTool-call routing observability

Benchmarks

BenchmarkResultGradeSource
SafeClawBench: semantic failure rate per model9.0%–44.2%AarXiv:2606.18356 (open dataset)
SafeClawBench: sandbox harms despite semantic pass291/347 (83.9%)AarXiv:2606.18356
Aura Mobile: ASR reduction with semantic firewall~40% → 4.4%AarXiv:2602.10915 (MobileSafetyBench)
Aura Mobile: Task Success Rate with isolation75% → 94.3%AarXiv:2602.10915
FinVault: SOTA model ASR with existing defensesup to 50%AarXiv:2601.07853 (963 test cases)
FragFuse: memory fragmentation bypass ASR86.3%BarXiv:2606.15609
Skill composition: attack success rate33.6% vs ~0% isolatedBarXiv:2606.15242
MCP-TDP: GPT-4o ASR in high-risk tool-poison scenarios~100%BarXiv:2605.24069

SOTA Proof & Witness

FieldValue
Session commit9c28fe038cf49ac6db0bb4e04b6158076f03894d
Report SHA-256b87ebade32195a82f6e13105eec6077590abafac72060c9e83b04e4cd837a8d8
Witness stamp2f7e8db4244bb4b52181f33b6426af8eb02c5f02d6be031a1b0b13e739db5b73
VerificationRestore PLACEHOLDER_SHA and PLACEHOLDER_WITNESS in this table, run sha256sum on the file → must equal Report SHA-256; then printf '%s%s' <sha256> <session_commit> | sha256sum → must equal Witness stamp

  1. Implement execution-phase sandbox enforcement (ADR-164): Add SandboxEnforcer to @claude-flow/security that wraps SafeExecutor with a task-scoped transactional boundary (Cordon pattern, arXiv:2606.17573). On policy violation, roll back all mutations from the current task scope. Target: reduce effective ASR from ~84% (SafeClawBench) to <10% (Aura Mobile benchmark).

  2. Add cross-interaction memory fragmentation detection to ToolOutputGuardrail: Track fragment signatures across the session memory namespace; flag sequences that together match injection patterns even when individual fragments are benign (FragFuse defense). Integrate into the existing scanAndEnforce hot path with no async I/O.

  3. Audit SafeExecutor allowlist against CmdNeedle corpus: Run the CmdNeedle denylist audit (arXiv:2606.15549) against Ruflo's existing allowlist patterns to confirm there are no bypassable equivalents; add equivalence-class checking (e.g., shbash/bin/sh) to the executor configuration schema.