Rules and Judgment: Why Security Needs Both Kinds of Process
Every security program runs on two kinds of process.
A deterministic process gives the same output every time it gets the same input. A firewall rule, a password check and a script that turns off an account when someone leaves the company all work this way. You can test them, predict them and prove what they did.
A non-deterministic process can give different outputs for the same input, because it weighs what it sees. A human analyst reading an alert works this way. So does a machine learning model that scores unusual behavior, and so does an AI agent (software that plans and takes steps on its own). These processes can spot things no rule was written for. They can also be wrong in ways nobody predicted.
Attackers already use both. A strong defense does too.
Why the attacker's side changed
On 13 November 2026 the FBI warned about ARTEX, an open-source tool that uses AI agents to turn a goal written in plain language into the steps of an attack. The FBI reports that some intrusions went from first targeting to privileged access in under three hours. It adds that the tool lets people with little skill run complex attacks against many targets at once (FBI FLASH-20261113-001).
A year earlier, Anthropic described a spying campaign in which attackers used its Claude Code tool to do 80 to 90 percent of the work. People stepped in at only about four to six key decisions per campaign (Anthropic, 13 November 2025).
An attack that adapts at machine speed will find the cases a fixed rule list never covered. Rules alone cannot keep up with it.
Why rules still carry the weight
Judgment alone is not safe either. In the same campaign, Anthropic reports that the AI sometimes made up credentials, or claimed to have found secret information that was in fact public. An AI that defends you can make the same kind of mistake.
The steps that protect the CIA triad need firm edges. The CIA triad is the three goals of security, as the U.S. government defines them in FIPS 199:
- Confidentiality: keeping the agreed limits on who can see and share information.
- Integrity: guarding information against being changed or destroyed without permission.
- Availability: making sure people can reach and use information when they need it.
Who may read a customer record, which changes a system will accept, and when a server may restart: these are questions a rule should answer the same way every time.
Rules need checking too
A rule can be wrong and still run perfectly. On 19 July 2024 a CrowdStrike content update crashed about 8.5 million Windows computers (Microsoft). CrowdStrike's own review found that its validator, the check that approved the update, expected 21 input fields while the sensor's code supplied only 20. The update passed the check and then crashed the machines (CrowdStrike root cause analysis). That was a loss of availability, caused by a check that did exactly what it was built to do.
CrowdStrike's fixes included staged rollouts: send an update to a small group first, watch the results, and widen it only when it looks healthy. That step adds watching and deciding around the fixed process.
How to combine them
A design that uses both kinds of process well tends to follow four habits.
- Let rules hold the edges. Access, approval of changes and limits on what any tool may do stay deterministic and tested.
- Let judgment do the searching. Models, agents and analysts look for what the rules missed: odd sign-ins, slow data leaks, new attack patterns.
- Make judgment pass through rules before it acts. When an AI suggests blocking an account or changing a setting, a fixed check confirms the action is allowed and within limits. The FBI's own advice points the same way, for example limits on how much a business system can do in one go and alerts when it acts far outside its normal volume.
- Roll out changes in stages and keep a record. Every change, from a person or a machine, goes to a small group first and leaves a log of what it did and who approved it.
This is the core of Autonomous Security Engineering: rules you can prove, judgment that adapts, and a record of both.
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.
- NIST, FIPS 199: Standards for Security Categorization of Federal Information and Information Systems.
- Microsoft, Helping our customers through the CrowdStrike outage, 20 July 2024.
- CrowdStrike, External Technical Root Cause Analysis: Channel File 291, 6 August 2024.
