Manifold security argues gateways fail AI agents

manifold.security1 min read

The author argues that traditional gateways fail to secure AI agents because the vast majority of agent activity consists of actions like tool calls, shell commands, and file edits rather than prompts and responses. According to the text, only eleven percent of agent events are prompts and responses, while tool calls make up fifty-one percent and agent-to-agent traffic accounts for twenty-five percent. Consequently, crucial operations occur at the local level and bypass boundary-based proxy filters entirely. Open article →