TaintGate v0.1.0

TaintGate v0.1.0

Initial release of provenance-aware runtime security for AI agents.

Untrusted data should not become authority.

Highlights

• Provenance-aware protected tool-call enforcement
• Deterministic ALLOW / REVIEW / BLOCK decisions
• Prompt-injection, secret, PII, shell, SQL, sensitive-path, and data-flow detection
• Human approvals and structured audit lifecycle
• Fail-closed structured argument validation
• Deterministic local attack simulator
• OpenAI Agents integration
• LangChain / LangGraph integration
• MCP runtime enforcement and returned-data provenance
• Retry-safe MCP post-execution provenance handling

Release validation

• Python 3.10–3.13
• 185 tests passing
• Ruff clean
• taintgate attack: 12 / 12 expected protections
• Wheel and sdist build verified
• Clean-wheel installation verified
• OpenAI Agents, LangChain/LangGraph, and MCP extras verified
• Final main CI green

Security scope

TaintGate is a runtime enforcement layer, not an OS sandbox or a replacement for framework/MCP authentication and authorization. See SECURITY.md and docs/THREAT_MODEL.md for boundaries and non-goals.