The goals remain. Speed and scale change.

Stealing data, gaining access and disrupting services remain attackers’ objectives. In its Annual Review 2026, the NCSC describes an acceleration of existing techniques. AI can assist operators; capabilities, results and the degree of autonomy vary with the context.

A convincing message can still be an attack

AI can help prepare and adapt social engineering messages. A well-written request that seems familiar and relevant to your work can still be fraudulent. Imagine a false request from a supplier: before acting, verify identity and authorisation through an established channel.

The window for intervention can shrink

The NCSC highlights how AI can accelerate vulnerability research and the development of tools to exploit weaknesses. Defenders may have less time to fix an exposed system. There is no universal deadline for every attack: it depends on the flaw, the environment and the attacker’s capabilities.

AI itself becomes a target

An agent that reads emails or documents and uses tools introduces additional areas to protect. Prompt injection can attempt to turn the content it reads into an instruction to execute. The impact depends largely on accessible data and permissions, so boundaries also need enforcement by connected systems.

Shorten the path from a signal to a fix

An alert is useful when someone can turn it into a decision. Identify affected systems, set priorities, prepare the update and verify the outcome: this is the patch management cycle described by NIST. An exposed service with a relevant flaw, for example, needs an owner, a decision on intervention and a subsequent check that it works as intended.

Where AgIPS fits

AgIPS proposes connecting monitoring, verification, action controls and remediation within rules set by people. The aim is to shorten the distance between a signal and a verified intervention. Trials will need to measure response times, issues identified and legitimate activities blocked by mistake. AI supports defence when its contribution can be controlled.

Further reading