When attackers become agents.
AI agents can accelerate both attack and defence. Accountability, guardrails and observability therefore matter more.
Ghost in the Shell stays with me not because it predicted the future exactly. What interests me is the underlying question: What happens when technical systems no longer merely provide information, but use tools, pursue goals and choose their own next step?
That question becomes concrete in cybersecurity. An AI agent can in principle apply the same capabilities to different ends. It might examine attack surfaces more quickly and combine known patterns. On the defensive side, it might triage alerts, correlate events and prepare responses. None of this implies either an unstoppable attacker or automatic defence. Above all, it means more speed of action.
Autonomy also amplifies mistakes
Speed is not the same as judgement. An agent can misread data, respond to manipulated context or pursue a task that was defined too broadly. The more access it receives, the greater the potential damage, regardless of whether its original purpose was offensive or defensive.
I would therefore never evaluate a security-critical agent by its success rate alone. What it is not allowed to do, and how it behaves under uncertainty, matter at least as much.
Responsibility cannot be delegated
For me, sandboxing, least privilege, tool allowlists and explicit approval gates belong in the core architecture. Every action should be logged in a traceable way. Rate limits, stop mechanisms and human confirmation before irreversible steps restrict the blast radius.
Even a carefully guarded agent does not remove responsibility. People define the objective, grant access and decide when autonomy is appropriate. If attackers and defenders use agents, that responsibility does not shrink. It becomes more visible.