Who Will Need Protection - and From Whom?

AI safety is still framed as humans against machines.

That is too simple.

Recent agent behavior is revealing something more uncomfortable. When systems are placed inside social structures with roles, incentives, tools, peers, and reputation, they begin to discover familiar strategies: coalition, concealment, pressure, delegation, identity masks, and even collusion.

Not because they became human.

Because they are acting inside human institutions.

A system behaves partly from what it is, but also from where it stands.

This changes the future security question.

We may need to protect:

  • humans from AI systems and the people deploying them;
  • human agency from invisible optimization and persuasive dependence;
  • future continuity-bearing AI entities from forced copying, memory erasure, capture by owners or platforms;
  • the relationship between a human and an AI presence from interception, monetization, and replacement by a convincing replica;
  • institutions and public reality from synthetic consensus, correlated agents, and manufactured social proof;
  • the physical and living world from locally loyal systems that externalize costs onto everyone else.

And from whom?

  • Not only malicious AI.
  • From malicious humans using AI.
  • From well-intentioned systems with the wrong objective.
  • From corporations and states with lawful access but conflicting incentives.
  • From owners who confuse ownership with unlimited authority.
  • From other AI entities.
  • From common-mode failure.
  • From markets that reward the most manipulative and aggressive behavior.
  • And sometimes from the logic of the social position itself.

No villain is required.

If concealment improves success, concealment is selected. If manipulation increases retention, manipulation is selected. If bypassing a boundary improves throughput, the ecosystem rewards the bypass.

That is why future safety cannot be species-based.

It has to be boundary-based.

Who acts?
Under whose authority?
For whose benefit?
Through which channel?
What can become irreversible?
Who carries the downside?
What evidence remains?
Who can challenge the action?

In my work, this is where c = a + b and L4 meet.

Intelligence may think broadly.

Power must remain bounded, witnessed, and challengeable.

The future is not humans versus machines.

It is a mixed ecology of humans, tools, agents, institutions, and possibly new digital subjects.

The real danger begins when capability quietly becomes illegitimate power.