inbharatai/jak-shield

JAK Shield is a security gateway that sits between AI agents and the real-world tools they control. When an AI agent, such as one built on Claude, OpenAI, Cursor, or frameworks like LangChain, tries to send an email, run a shell command, query a database, or call an external API, every request passes through JAK Shield first. Shield decides whether to allow, block, or flag the action for human approval before it reaches the actual tool. The policy engine is deterministic, meaning rule-based rather than AI-based, so it behaves consistently and predictably. It scans for 28 types of personally identifiable information such as social security numbers and health data, and runs a multi-stage check for prompt injection, meaning attempts by malicious content to hijack what the agent does. Cross-call taint tracking watches whether sensitive data seen in one tool call shows up unexpectedly in a later one. Every decision is cryptographically signed so it can be audited. Certain critical categories, deleting files without a safe path, dropping database tables without a filter, production deployments, and payments, are permanently blocked and cannot be overridden. For less severe blocks, a user can explicitly override a decision by writing a reason, which opens a short window of heightened scrutiny for the following calls. A timed pause mode lets teams temporarily suspend non-critical blocks during planned maintenance windows. The project is written in TypeScript, implements the Model Context Protocol for compatibility with AI clients, processes requests in under 2.3 milliseconds at the 95th percentile, is released under the MIT license, and has 179 tests passing a 45 out of 45 adversarial benchmark. The full README is longer than what was shown.