What is the difference between agent governance and AI security?

AI security defends the model and the pipeline around it. Agent governance decides what an agent may do with the authority it holds, and proves what it did. Both are needed; they answer different questions, and conflating them leaves the second one unanswered.

The two terms are used interchangeably by buyers and by vendors, and the confusion is costly, because a team that has bought one believes it has the other. The cleanest separation is by the question each answers.

Two questions

• AI security asks: can the model or the system around it be attacked, poisoned, jailbroken, extracted or made to leak? Its subject is the intelligence and its inputs. Its tools are red teaming, prompt-injection defences, model provenance, data-pipeline integrity.
• Agent governance asks: given an agent that acts, was this action authorized for this identity under this delegation at this moment, can the record of it be proven to a third party, and could a human have stopped it? Its subject is authority and consequence. Its tools are an authorization gate before the act, a tamper-evident record after it, behavioural monitoring, and a fleet-wide halt.

Models reason. Agents act. Protocols transport. Identity systems identify. Payment networks settle. Agentomy governs authority and consequence: it owns authority, not execution. A perfectly secured model can still be told, legitimately, to do the wrong thing; governance is the layer that decides whether the action may happen and proves what happened either way.

Where they meet

They meet at the tool call. An agent's consequential acts arrive as tool calls, API calls and transactions, and that is where an interposing governance layer sits: in front of the server, deciding each call before it runs and failing closed when it cannot decide. Prompt-injection detection is one input to that decision; it is not a substitute for it. Governance stays useful as models improve, because a smarter model still needs its authority bounded and its actions recorded.

Security hardens the actor. Governance bounds the authority and proves the act. A regulated deployment needs the second in writing.

Where to verify this yourself

• The Agent Governance Protocol, the open specification for the primitives above: agentomy.com/agp.
• The MCP gateway that interposes on tool calls and its capability profile: agentomy.com/mcp-gateway and agentomy.com/adapters.json.
• Core concepts, including the question the layer answers on every action: agentomy.com/docs/concepts.