Phinq

Product Introduction

1. Definition: Phinq is an open-source runtime governance and security layer specifically engineered for AI agents and autonomous systems. It functions as a middleware proxy or SDK that intercepts, evaluates, and enforces policy on every tool call an AI agent attempts to make before execution.
2. Core Value Proposition: Phinq exists to provide structural, enforceable safety for AI agents in production. It prevents irreversible or high-risk AI actions—like data deletion, financial transactions, or sending external communications—from executing without explicit human oversight, thereby mitigating AI agent risks and enabling secure AI agent deployment.

Main Features

1. Runtime Interception & Tool Call Inspection: Phinq operates at the execution boundary, intercepting every API call an AI agent makes to external tools or functions. It inspects the tool name, arguments, and relevant session context in real-time before the call is forwarded to the target service.
2. Risk-Based Action Classification & Policy Enforcement: The system classifies each intercepted action based on configurable risk policies. Actions deemed "safe" or reversible (e.g., database reads) are allowed to pass through automatically. Actions classified as "risky" or "irreversible" (e.g., delete_records, send_payment) are automatically held in a pending state, triggering a human-in-the-loop approval workflow via a configured notification channel (e.g., Slack, email).
3. Tamper-Evident, Hash-Chained Audit Logging: Every decision—pass, hold, approve, or deny—is immutably recorded in an append-only JSONL audit log. Each log entry contains a cryptographic hash of the previous entry, creating a verifiable chain of custody. This provides a forensic audit trail for compliance, security incident response, and demonstrating AI governance.

Problems Solved

1. Pain Point: The "sleeping developer" problem, where an autonomous AI agent can execute damaging, irreversible actions without a pre-execution decision point. Traditional governance via prompts or guidelines is advisory and can be ignored by the agent.
2. Target Audience: AI/ML Engineers and DevOps Teams building production AI agent systems; Security and Compliance Officers (SecOps, GRC) requiring audit trails for AI actions; Product Teams integrating AI agents into customer-facing applications where safety is critical.
3. Use Cases: Governing customer support agents that can issue refunds or delete account data; overseeing AI coding assistants with access to production deployment tools; monitoring AI sales agents that can send emails or update CRM records; ensuring compliance for financial or legal AI agents handling sensitive operations.

Unique Advantages

1. Differentiation: Unlike prompt-based safety or "guardrail" skills that an agent can circumvent, Phinq provides structural enforcement by sitting at the API call layer. It is also distinct from static code analysis or pre-deployment testing, as it governs behavior dynamically at runtime.
2. Key Innovation: The combination of runtime interception with a fail-closed human approval gate for risky actions, backed by a cryptographically verifiable audit log. This creates a closed-loop governance system where oversight is mandatory for critical actions and all decisions are provably recorded.

Frequently Asked Questions (FAQ)

1. How does Phinq prevent an AI agent from causing damage? Phinq acts as a mandatory gateway for all agent tool calls. It uses runtime interception to pause high-risk actions like deletions or payments, requiring explicit human approval before they proceed. If no decision is made, the system fails closed, blocking the action by default.
2. Is Phinq compatible with LangChain, OpenAI Assistants, or Claude API? Yes. Phinq is framework-agnostic. It offers multiple integration boundaries: a forward proxy for HTTP-based agents (OpenAI, Anthropic, Gemini), an in-process SDK, and a plain HTTP gate. This makes it compatible with LangChain, CrewAI, AutoGen, and any system using standard LLM provider APIs.
3. Does using Phinq add significant latency to AI agent responses? Minimal latency is added for safe, pre-approved actions as they pass through directly. Latency is only introduced for actions classified as risky, which wait for human review. The performance cost is incurred precisely where oversight is required, not on every query.
4. Can the Phinq audit log be tampered with or falsified? No. The audit log is designed to be tamper-evident. It uses a hash chain where each entry's hash depends on the previous entry's content and hash. Any alteration to a historical log entry would break the cryptographic chain, making manipulation immediately detectable upon verification.
5. What happens if the Phinq service or approval channel goes down? The system is designed for resilience. Configuration determines the fail-safe behavior. Typically, it can be set to "fail closed" (block all unvetted actions) or "fail open" (allow actions after a timeout) based on the risk profile of the deployment, ensuring operational continuity is managed according to policy.