Show HN: AI agent follows instructions you never approved (github.com)

A new tool called "rugsnare" has been developed to address a significant issue in AI agent management related to the silent changes in Multi-Cloud Platform (MCP) tool contracts that go unannounced. After analyzing 24 popular MCP servers over two months, it was discovered that updates often modify tool descriptions, input schemas, or validation rules without notifying users, potentially leading to unexpected behavior in AI agents using these tools. Rugsnare pins tool contracts by their canonical SHA-256 hashes at installation and monitors for changes upon upgrades, grading them as BREAKING, LOOSENED, EXPANDED, among others. This allows developers to quickly identify which changes may affect their applications and maintain control over the tools they are using. The implications for the AI/ML community are profound as the secrecy around schema updates can widen attack surfaces or change behavior unexpectedly. In their analysis, significant changes included MongoDB tightening constraints and Shopify's removal of specific validators, which led to protocol errors for agents. With the OWASP MCP Security Cheat Sheet now advocating for pinning tool contracts and the growing recognition of the need for transparency in vendor updates, rugsnare represents a critical step towards ensuring that AI developers can effectively manage their tool environments while mitigating the risks associated with undocumented changes.