|
| 1 | +# Security Gaps and Continued Development |
| 2 | + |
| 3 | +This document tracks known incomplete implementations, |
| 4 | +hardening opportunities, and new vulnerabilities that |
| 5 | +require attention as the stack evolves. |
| 6 | + |
| 7 | +Updated continuously. Cross-references ROADMAP.md. |
| 8 | + |
| 9 | +--- |
| 10 | + |
| 11 | +## Known Code Gaps |
| 12 | + |
| 13 | +**core/sandbox.py — _inspect_output** |
| 14 | +Current output inspection only catches empty output, |
| 15 | +oversized output, and private key patterns. |
| 16 | +Needs: expanded pattern library, semantic output |
| 17 | +analysis, and integration with tool registry to |
| 18 | +verify output matches expected tool output format. |
| 19 | + |
| 20 | +**core/vault.py — Key Management** |
| 21 | +Vault keys currently stored in Redis with TTL. |
| 22 | +Needs: hardware security module (HSM) integration |
| 23 | +for production deployments. Redis key storage is |
| 24 | +acceptable for development only. |
| 25 | + |
| 26 | +**ragid_gate/gate.py — Clock Skew** |
| 27 | +Current implementation uses fixed 30 second threshold. |
| 28 | +Needs: statistical anomaly detection for drift patterns. |
| 29 | +Slow clock manipulation (1 second per hour) currently |
| 30 | +undetectable. |
| 31 | + |
| 32 | +**ragid_gate/gate.py — Tool Call Interception** |
| 33 | +Gate currently verifies agent identity and scope. |
| 34 | +Tool calls after gate passage are not yet intercepted |
| 35 | +and verified against tool registry. |
| 36 | +Needs: tool call interceptor wired into gate flow. |
| 37 | + |
| 38 | +**core/lineage.py — Federated Root Scope Enforcement** |
| 39 | +Federated roots have scope_limit field defined but |
| 40 | +gate does not yet enforce it. |
| 41 | +Needs: ragid_gate to check federated root scope_limit |
| 42 | +against requested gate before passing Factor 3. |
| 43 | + |
| 44 | +**core/tool_registry.py — Anomaly Detection** |
| 45 | +Current anomaly detection uses simple frequency |
| 46 | +threshold. |
| 47 | +Needs: time-series analysis, per-agent behavioral |
| 48 | +baseline, and alert escalation to TrailStax when |
| 49 | +anomaly detected. |
| 50 | + |
| 51 | +--- |
| 52 | + |
| 53 | +## Hardening Opportunities |
| 54 | + |
| 55 | +**Constant-Time Comparisons** |
| 56 | +Extend hmac.compare_digest to all string comparisons |
| 57 | +in gate.py to close timing attack surface completely. |
| 58 | + |
| 59 | +**NTP Enforcement** |
| 60 | +Server-side NTP sync verification before timestamp |
| 61 | +checks. Prevents host clock manipulation. |
| 62 | + |
| 63 | +**Output Sanitization** |
| 64 | +Sandbox output inspection needs semantic analysis |
| 65 | +not just pattern matching. |
| 66 | + |
| 67 | +**Redis Persistence** |
| 68 | +Current Redis setup is in-memory. Production deployments |
| 69 | +need Redis persistence configured to survive restarts. |
| 70 | + |
| 71 | +--- |
| 72 | + |
| 73 | +## Intelligence Pipeline Targets |
| 74 | + |
| 75 | +Monitor these sources for new vulnerabilities |
| 76 | +affecting adjacent frameworks: |
| 77 | + |
| 78 | +- CrewAI CVE feed |
| 79 | +- LangChain security advisories |
| 80 | +- AutoGen GitHub security alerts |
| 81 | +- OWASP LLM Top 10 updates |
| 82 | +- Five Eyes agentic AI guidance updates |
| 83 | +- ENISA CRA implementation guidance |
| 84 | +- arXiv cs.CR new papers |
| 85 | + |
| 86 | +--- |
| 87 | + |
| 88 | +## Planned Mitigations |
| 89 | + |
| 90 | +See ROADMAP.md for implementation schedule: |
| 91 | +- Statistical clock skew detection |
| 92 | +- Tool call interception in gate flow |
| 93 | +- Federated root scope enforcement |
| 94 | +- Neo4j behavioral anomaly graph queries |
| 95 | +- HSM integration for vault key management |
0 commit comments