Why Security Still Needs a Human in the Loop
A human in the loop is a person who must look at an action and say yes before a system takes it. In security, that action might be locking a user out, shutting down a server, deleting data or pushing an update to every computer in the company.
Automation now does much of the work in a security team, and AI agents (software that plans and takes steps on its own) do more of it every month. That is good news for speed. It also raises a hard question: which steps should a machine never take alone?
Attackers have already cut people out
On 13 November 2026 the FBI warned about ARTEX, an open-source tool that uses AI agents to plan and carry out attacks. The FBI says it needs little human input beyond setting the goal. Some intrusions reached privileged access in under three hours (FBI FLASH-20261113-001).
In November 2025, Anthropic reported a spying campaign in which AI did 80 to 90 percent of the work. People stepped in at only about four to six key decisions per campaign. At its peak the AI sent thousands of requests, often several per second (Anthropic).
Defenders need automation to keep up. The same report shows where a person still matters most. The AI sometimes made up credentials, or claimed to have found secret information that was in fact public. A defender's AI can make the same kinds of mistakes. When it does, someone has to catch them before they become actions.
What can go wrong without a person
The OWASP Gen AI Security Project lists excessive agency in its Top 10 risks for AI applications built on large language models (the kind of AI behind chat assistants). Excessive agency means an AI system has enough power to do damage when its output is unexpected, unclear or manipulated by an attacker. One of OWASP's fixes is to "require a human to approve high-impact actions before they are taken" (OWASP LLM06:2025).
The mistake does not have to come from AI. On 19 July 2024 a routine CrowdStrike content update crashed about 8.5 million Windows computers (Microsoft). Among CrowdStrike's fixes were staged rollouts, with each wider group of computers reached only after the results were watched. It also gave customers control over when these updates reach their own machines (CrowdStrike root cause analysis). Both put a pause, and a chance for a person to decide, before a change reaches everything.
What the rule books say
- The EU AI Act, Article 14, says high-risk AI systems must be built so people can oversee them. Those people must be able to understand what the system can and cannot do, stay alert to the habit of trusting its output too much, and step in or stop it (Article 14).
- The U.S. NIST AI Risk Management Framework says AI systems can run anywhere from fully on their own to fully by hand. It says the roles of the people who oversee them need to be clearly defined (NIST AI 100-1).
How a human in the loop protects the CIA triad
The CIA triad is the three goals of security, as defined in the U.S. government's FIPS 199: confidentiality (keeping the limits on who can see information), integrity (guarding information against changes without permission) and availability (making sure people can reach information when they need it). A person approving the right steps protects each one.
- Confidentiality. A person approves any action that shares data outside the company or widens who can see it.
- Integrity. A person approves changes to records, settings and permissions that cannot be easily undone.
- Availability. A person approves shutting down systems or rolling an update out to everyone, the kind of step that can stop the business.
Where to put the person
A person who must approve everything will soon approve everything without reading it. The design work is choosing the few steps that matter.
- Sort actions by impact. Routine, reversible steps run on their own. High-impact or permanent steps wait for a yes.
- Show the person what they need to decide. The approval request says what will happen, why, and what it touches.
- Keep a stop button. Anyone responsible can pause the automation at once.
- Record every decision. Each action and each approval is logged, so the team can show later who said yes and why.
This is how Autonomous Security Engineering works: machines handle the volume and the speed, and people make the calls that can hurt the business most.
Read more about Autonomous Security Engineering, or schedule a discovery call.
Sources
- FBI, FLASH-20261113-001: Threat Overview of ARTEX Autonomous Penetration Testing Framework, 13 November 2026.
- Anthropic, Disrupting an AI-orchestrated cyber espionage campaign, 13 November 2025.
- OWASP Gen AI Security Project, LLM06:2025 Excessive Agency.
- Microsoft, Helping our customers through the CrowdStrike outage, 20 July 2024.
- CrowdStrike, External Technical Root Cause Analysis: Channel File 291, 6 August 2024.
- EU Artificial Intelligence Act, Article 14: Human Oversight.
- NIST, AI 100-1: Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023.
- NIST, FIPS 199: Standards for Security Categorization of Federal Information and Information Systems.
