AgentArmor

Description

Open-source 8-layer security framework for AI agents (pip install agentarmor-core). Maps to all 10 OWASP ASI risks (December 2025 spec). Layers: L1 Ingestion (prompt injection, Unicode stego), L2 Storage (AES-256-GCM + BLAKE3), L3 Context (parameterized context + canary tokens), L4 Planning (risk scoring READ=1 through ADMIN=10, chain depth limits), L5 Execution (egress rate limiting, approval gates), L6 Output (Presidio PII redaction), L7 Inter-agent (HMAC-SHA256 mutual auth), L8 Identity (JIT permissions). Known limitation: L4 scores by verb only (SELECT * WHERE 1=1 scores same as a targeted lookup).

Key claims

Relations

Sources

src-20260419-be936b1b7d4c