Agent Cyber Sentinel is an open-source Python security layer for autonomous AI agents and Model Context Protocol (MCP) servers. Its stated purpose is to screen untrusted prompts and retrieved context, prevent leaked credentials from reaching prompts or logs, and stop unsafe tool activity before execution. The project describes three main threat classes: indirect prompt injection and jailbreak-style instruction hijacking, exposure of API keys and other secrets, and SSRF or destructive commands issued through HTTP and code-execution tools. Its runtime pipeline applies a prompt-injection firewall, entropy and pattern-based secret redaction, a tool-call sanitizer, and output screening. It says blocked requests can raise dedicated errors, while permitted or blocked actions produce POD-9 Proof-of-Defense envelopes with deterministic SHA-256 audit seals. The claimed overhead is below 1 millisecond, although the supplied material does not provide independent benchmark details. Supported detections include common OpenAI, Anthropic, AWS, Google, GitHub, JWT, private-key, database-URI, cloud-metadata, private-network, shell, SQL, and path-traversal patterns. The repository includes MCP tools, an OpenAI client wrapper, LangChain/LangGraph and CrewAI integrations, and a terminal CLI. Installation is documented through pip, and the project is released under Apache 2.0.
