AI can respond at machine speed. That promise matters because security teams already face attacks that move faster than human approval chains. Technology can detect something in seconds, understand what is happening, and recommend the appropriate response, but someone still has to approve what happens next.
Sebastiaan Bäck, Field CISO at Halcyon, described the opportunity plainly: “AI potentially gives us something security teams have wanted for decades: the ability to identify, understand, and respond to an attack at machine speed.” The machinery is not the only obstacle, though. Many organisations are deploying AI into the same operating model they had before, which leaves the central question untouched: who has permission to act?
Detection Is Not the Same as Response
If nobody gives a system a mandate to respond while an attack is happening, the response waits until after the incident. That delay turns a technical capability into another recommendation sitting in an approval queue — a familiar place for important decisions to go and quietly lose the race.
Most organisations have not properly solved response authority for their human security teams, either. Decisions such as isolating a machine, taking a server offline, or interrupting a production process are often unclear or negotiated during an incident, when the cost of hesitation is already rising.
AI exposes how unsustainable that model has become. A system that can identify and understand an attack at machine speed cannot deliver its value if every action still depends on improvised permission, unclear ownership, or an argument over whether a server may leave the network.
The problem is not solved by telling security systems to “be autonomous.” That phrase skips the part where an organisation decides what autonomy means, which actions are allowed, and where independent control must remain in place. Without those decisions, AI can recommend an appropriate response while the incident continues without it.
Permission Requires Context and Controls
Giving AI permission to act creates another requirement: asset intelligence and business context. A system needs to understand what it is affecting before it isolates a machine, takes a server offline, or interrupts a production process. Authority without context is not governance; it is a faster route to an unclear outcome.
Boards and executive teams need to settle those boundaries before an attack occurs. Bäck put the responsibility directly on them: “Boards and executive teams need to decide what authority they are prepared to give their security systems before the attack happens, not while everybody is watching it unfold.”
That decision also needs controls independent of the system taking action. Alex Salazar, Co-Founder and CEO of Arcade.dev, offered a useful comparison: “Applications don’t grant their own permissions without independent sign-off. Traders don’t approve their own trades without pre-authorization standards.” Every function has an independent control sitting to the side, and AI agents require the same seriousness.
Yet AI agent governance is a different discipline than governing people. An organisation must account for the system’s access, the business context behind each decision, and the authority attached to each response. Treating an AI agent like another employee misses the structure of the problem — and employees do not usually make decisions at machine speed.
The practical task is clear: define response authority before the incident, connect that authority to accurate asset intelligence and business context, and place independent controls beside the system. AI can identify, understand, and respond to attacks at machine speed. Without permission to act, it remains an exceptionally fast observer.
Based on
