OWASP Top 10 for
Agentic Applications

OWASP published the Top 10 for Agentic Applications in December 2025 (ASI01–ASI10). MAREF is a governance-first runtime built to answer these risks with runtime-enforced controls, not policy documents.

Full engineering mapping with key-code and TLA+ correspondence: owasp-agentic-top10-mapping.md.

Risk MAREF control Evidence location
ASI01 — Agent Goal HijackSubgoal interceptor + 4-level safety decision treesrc/maref/subgoal/
ASI02 — Tool MisuseTool-Gateway chokepoint, per-tool least privilegesrc/maref/tools/registry.py
ASI03 — Identity & Privilege AbuseZero-trust per-agent Ed25519 identity, time-scoped credentialssrc/maref/identity/
ASI04 — Agentic Supply ChainTool/plugin provenance checks; MCP & A2A transport hardeningsrc/maref/mcp/
ASI05 — Unexpected Code ExecutionSandbox isolation + circuit breaker HALTsrc/maref/governance/circuit_breaker.py
ASI06 — Memory & Context PoisoningTrust Engine v2 drift detection (KL/JS/Hellinger)src/maref/evaluation/
ASI07 — Insecure Inter-Agent CommsSigned inter-agent messages; per-agent identity isolationsrc/maref/security/
ASI08 — Cascading FailuresGray Code FSM — Hamming distance = 1 on every transition; HALT absorbing statesrc/formal/MarefJoint34.tla
ASI09 — Human-Agent Trust ExploitationMode/SafetyGate escalation for high-risk user-influenced actionssrc/maref/governance/
ASI10 — Rogue AgentsAutonomy tiering + behavioral drift → auto policy lockdownsrc/maref/governance/policy.py

Evidence, not slogans

Every row above links to source code or a test directory. Two headline measures of "does this actually run?": the 10-risk mapping is CI-verifiable, and the formal specs pass TLC model checking via pytest tests/formal/.

Related reading: MAREF's 10/10 coverage explained · NIST alignment · For AI Agents / LLMs