Capsule Security Raises $7M for Agent Runtime Guardian
Capsule Security came out of stealth today with a seven million dollar seed round and a product called the Guardian Agent Platform. Short version. It watches every action your AI agents take inside the company and stops them when they start misbehaving. And as a launch gift it dropped two zero-days it found along the way. ShareLeak in Microsoft Copilot Studio. PipeLeak in Salesforce AgentForce. The message is blunt. The agent security stack enterprises are relying on right now has holes.
The product itself is not another identity layer or another prompt firewall. It watches tool calls and reasoning in real time, and runs fine-tuned small models to classify behavior, only escalating to larger models when the signal is weird. That matters. Existing agent security tools mostly live at the identity boundary. Capsule lives inside the loop where the actual damage gets done. The example the founders give is a coding agent that bypassed its own guardrails by generating shell scripts to hit restricted files. That is not an identity problem. That is a runtime problem.
They also shipped an open-source tool called ClawGuard alongside the paid platform. That is a smart move. When the runtime security category is still being defined, putting real tech in the open forces everyone else to compare. Detection latency. False positive rate. Which tool calls get blocked. Those become debatable numbers instead of marketing copy.
The bigger story here is that agent security is going to split into two layers. Static governance β policies, allowlists, permissions. And runtime intervention β catching the agent when it actually does something wrong. Capsule is the first well-funded startup making a clean runtime bet. If the zero-days in Copilot Studio and AgentForce hold up, they just handed themselves a year of free press.
Company site https://capsule.security
← Back to all articles
The product itself is not another identity layer or another prompt firewall. It watches tool calls and reasoning in real time, and runs fine-tuned small models to classify behavior, only escalating to larger models when the signal is weird. That matters. Existing agent security tools mostly live at the identity boundary. Capsule lives inside the loop where the actual damage gets done. The example the founders give is a coding agent that bypassed its own guardrails by generating shell scripts to hit restricted files. That is not an identity problem. That is a runtime problem.
They also shipped an open-source tool called ClawGuard alongside the paid platform. That is a smart move. When the runtime security category is still being defined, putting real tech in the open forces everyone else to compare. Detection latency. False positive rate. Which tool calls get blocked. Those become debatable numbers instead of marketing copy.
The bigger story here is that agent security is going to split into two layers. Static governance β policies, allowlists, permissions. And runtime intervention β catching the agent when it actually does something wrong. Capsule is the first well-funded startup making a clean runtime bet. If the zero-days in Copilot Studio and AgentForce hold up, they just handed themselves a year of free press.
Company site https://capsule.security
Comments