At a recent security conference in Washington, physical security leaders, cyber defenders, AI policy specialists, and counterintelligence professionals sat at the same table. That alone said something. Five years ago those groups attended different conferences and spoke different languages.
They are now defending the same terrain. They still do not share a framework for it.
Four shifts that matter to boards
The hardest deception is built from true facts
The conversation has moved from detecting bad actors to detecting bad patterns. The most effective fraud no longer relies on false information. It arranges true information to mislead. Increasingly, the system doing the arranging is an AI.
Social engineering got cheap
Operators are less worried about malware than about AI-powered deception. A cloned voice authorizing a wire transfer. A phishing email written in the target's own style. The cost of a convincing lie has collapsed, and every control that depended on a lie being expensive has weakened with it.
The insider no longer needs a motive
Insider threat programs are built on a human model: a person with access, a grievance, and an opportunity. The next insider may be an AI agent with real permissions, standing access to production systems, and no one watching it. It can do insider-scale damage with no intent at all. The root cause is a governance gap: no one decided who owns the agent's access.
The pause is gone
Human decision cycles used to include time for someone to notice an error. AI recommends and acts almost instantly. Oversight that happens after the fact now happens after the damage. Boards that review AI governance once a year as a policy document are governing a decision cycle that no longer exists.
Where this sits in the Four Signals Framework
This is an Exposure problem. The Exposure signal asks what data, systems, and decision points AI touches, and who can see them. In a converged threat environment, the honest answer at most organizations is that no single function can see all of it. Physical security sees the facility. The CISO sees the network. Legal sees the policy. The AI agent with standing access sits across all three and reports to none of them.
That is why The Reed Group works at the boundary between physical and digital risk. The exposure lives in the seams between functions.
Questions for the board
- Do we have an inventory of every AI system and agent with standing access to our systems or data?
- For each one, who owns its permissions, and who would notice if it acted on bad information?
- Which of our controls assume that a convincing deception is expensive to produce?
Inventory every AI system and agent with standing access. For each one, ask how long it would take to notice if it acted on bad information with no human involved. If the answer makes you uncomfortable, you are looking at an unmanaged insider.
An earlier version of this piece appeared in Northern Signal, Tim Reed's briefing on AI governance and security.