Google opened early access to an MCP server for Google Home, allowing third-party AI agents to interact with smart-home devices and access home event history. It marks one of MCP's first moves into controlling physical environments rather than software tools. Practitioners are weighing permission boundaries for agents whose tool calls carry physical consequences.
Bouncer is a local MCP proxy that gates outbound tool calls from AI agents, denying calls to destinations derived from untrusted content (like web pages or emails) and allowing only explicitly trusted ones. It uses deterministic Python policy enforcement rather than an LLM in the enforcement path, preventing prompt-injection attacks that trick agents into leaking API keys or other sensitive data. It solves the problem of agents acting on malicious instructions embedded in untrusted content, protecting users and developers building agentic systems.
Bouncer is a local MCP proxy that gates outbound tool calls from AI agents, denying calls to destinations derived from untrusted content (like web pages or emails) and allowing only explicitly trusted ones. It uses deterministic Python policy enforcement rather than an LLM in the enforcement path, preventing prompt-injection attacks that trick agents into leaking API keys or other sensitive data. It solves the problem of agents acting on malicious instructions embedded in untrusted content, protecting users and developers building agentic systems.
An Agent Proxy that sits between AI agents and MCP servers to control what tools and API resources agents can access. It keeps API keys out of agent context, restricts hosts, methods, and paths, and filters which MCP tools are exposed. Aimed at developers running agents like Codex who need policy enforcement and auditing.
An Agent Proxy that sits between AI agents and MCP servers to control what tools and API resources agents can access. It keeps API keys out of agent context, restricts hosts, methods, and paths, and filters which MCP tools are exposed. Aimed at developers running agents like Codex who need policy enforcement and auditing.
A new tool intercepts MCP (Model Context Protocol) traffic in real time to block agents from reading .env files and executing dangerous commands. It addresses security risks in agentic AI workflows by enforcing policy at the protocol level.
A new tool intercepts MCP (Model Context Protocol) traffic in real time to block agents from reading .env files and executing dangerous commands. It addresses security risks in agentic AI workflows by enforcing policy at the protocol level.
A developer on Reddit shared that they gave internal AI agents broad production API keys to speed up demos, and never tightened permissions afterward. One agent now has read/write access to a production database with minimal audit trail, highlighting common security pitfalls in agent deployments.
A developer on Reddit shared that they gave internal AI agents broad production API keys to speed up demos, and never tightened permissions afterward. One agent now has read/write access to a production database with minimal audit trail, highlighting common security pitfalls in agent deployments.
Two Medium posts argue that natural-language instructions like "don't touch production" are not a security boundary for AI agents, and that builders should define narrowly scoped permissions before deployment. The thread offers practitioners guidance on least-privilege design for agent systems.
Two Medium posts argue that natural-language instructions like "don't touch production" are not a security boundary for AI agents, and that builders should define narrowly scoped permissions before deployment. The thread offers practitioners guidance on least-privilege design for agent systems.