# ProvnAI

> ProvnAI builds the security and governance layer for autonomous AI agents. McpVanguard is the open-source MCP security surface available today, while VEX Protocol reflects the governed-execution and evidence architecture being developed through pilot and design-partner work. The focus is the execution boundary: the layer where AI model output becomes privileged real-world action.

## Products

- [McpVanguard](https://www.provnai.com/mcp-vanguard): Open-source security proxy for the Model Context Protocol (MCP). Intercepts and inspects routed tool calls before execution and can block configured metadata-poisoning, tool-payload, SSRF-style, path-traversal, and tool-call-hijacking patterns.
- [VEX Protocol](https://www.provnai.com/vex-protocol): Governed-execution and evidence architecture for verifiable agent actions. Pilot-oriented surface for teams exploring higher-assurance control and review.

## Solutions

- [Use Cases](https://www.provnai.com/use-cases): How security teams, platform engineers, and governance teams deploy ProvnAI infrastructure.
- [Compliance](https://www.provnai.com/compliance): Evidence-forward support for regulated environments, EU AI Act, DORA, and SOC2 alignment.

## Research

- [Layered MCP Runtime Enforcement](https://www.provnai.com/blog/layered-mcp-runtime-enforcement): McpVanguard v2.1.0 release analysis explaining preflight normalization, deterministic policy, camouflage detection, semantic advisory scoring, behavioral risk, and explicit policy composition.
- [MCP Prompt Injection: Tool-Calling Vulnerabilities and Defenses](https://www.provnai.com/research/prompt-injection): How prompt injection reaches MCP tool calls through user input, RAG, and tool results, plus defenses with deterministic policy and SSRF blocking.
- [Adversarial RAG and Context Poisoning for AI Agents](https://www.provnai.com/research/adversarial-rag): Analysis of poisoned retrieval content, indirect prompt injection in AI agents, and MCP boundary controls for adversarial RAG risk.
- [Deterministic MCP Proxy: Policy Enforcement Outside the Model](https://www.provnai.com/research/deterministic-proxy): How McpVanguard enforces policy at the MCP boundary without trusting model judgment as the enforcement layer.

## AI Security Glossary

- [Prompt Injection](https://www.provnai.com/glossary/prompt-injection): Foundational LLM vulnerability involving instruction override.
- [Indirect Prompt Injection](https://www.provnai.com/glossary/indirect-prompt-injection): Adversarial instructions embedded in retrieved external content.
- [Context Poisoning](https://www.provnai.com/glossary/context-poisoning): Persistence-oriented attacks against agent memory and beliefs.
- [Tool-Call Hijacking](https://www.provnai.com/glossary/tool-call-hijacking): Unauthorized invocation of sensitive MCP tool capabilities.
- [SSRF (AI Context)](https://www.provnai.com/glossary/ssrf): Agent-driven exploitation of cloud metadata and internal services.
- [Data Exfiltration](https://www.provnai.com/glossary/data-exfiltration): Covert transmission of context data via outbound tool parameters.
- [LLM Jailbreak](https://www.provnai.com/glossary/llm-jailbreak): Bypassing model-layer behavioral constraints via adversarial engineering.
- [Path Traversal](https://www.provnai.com/glossary/path-traversal): Unauthorized filesystem access via tool parameter manipulation.
- [Privilege Escalation](https://www.provnai.com/glossary/privilege-escalation): Acquisition of unauthorized capabilities via semantic drift.
- [Permission Drift](https://www.provnai.com/glossary/permission-drift): Incremental expansion of agent scope beyond initial intent.
- [Evidence Capsule](https://www.provnai.com/glossary/evidence-capsule): Cryptographic evidence construct preserving the governance context of governed execution.
- [Deterministic Proxy](https://www.provnai.com/glossary/deterministic-proxy): Binary, rule-based enforcement at the execution proxy boundary.
- [Merkle Audit Trail](https://www.provnai.com/glossary/merkle-audit-trail): Tamper-evident, append-only cryptographic logging for agent actions.
- [Execution Boundary](https://www.provnai.com/glossary/execution-boundary): Formally defined perimeter for authorized agent actions.
- [Authority Control](https://www.provnai.com/glossary/authority-control): Binding actions to verified principals and declared intent scopes.
- [Intent Attestation](https://www.provnai.com/glossary/intent-attestation): Binding of agentic sessions to authorized goals.
- [TEE Isolation](https://www.provnai.com/glossary/tee-isolation): Hardware-enforced runtime isolation being explored within the architecture for pilot and research use.
- [MCP Tool Manifest](https://www.provnai.com/glossary/mcp-tool-manifest): Structured declaration of available tools and their capability metadata.
- [Per-Action Verification](https://www.provnai.com/glossary/zero-trust-agent-model): Per-action verification architecture for autonomous AI systems.
- [Witness Log](https://www.provnai.com/glossary/witness-log): Tamper-evident execution record preserving what crossed the boundary.
- [MCP Standard](https://www.provnai.com/glossary/mcp): The open standard for AI model-resource interaction (Anthropic).
- [VEX Protocol (Glossary)](https://www.provnai.com/glossary/vex): The execution-boundary governance layer for consequential AI actions.

## Architecture & Concepts

- [Manifesto](https://www.provnai.com/manifesto): The thesis behind governed execution — why autonomous AI systems require a verifiable execution layer.
- [Glossary Index](https://www.provnai.com/glossary): Full AI security knowledge base — 23 terms across threat vectors and architecture primitives.

## Company

- [About](https://www.provnai.com/about): Mission, history, and institutional context for ProvnAI.
- [Contact](https://www.provnai.com/contact): Reach the ProvnAI team for pilots, architecture collaboration, or security disclosures.
- [Security](https://www.provnai.com/security): Responsible disclosure policy.

## Open Source

- GitHub: https://github.com/provnai
- PyPI (McpVanguard): https://pypi.org/project/mcp-vanguard/