Lower Barriers, Faster Attacks: How AI Changes the Threat Landscape

Cyberattacks depend on practical knowledge: understanding an unfamiliar system, selecting a technique, adapting a tool and interpreting the result. Each step creates friction. An operator has to find the right information and turn it into something that works in the environment in front of them.

AI can reduce that friction. It can explain technical material, help adapt code and interpret feedback within the same conversation. When connected to tools, it can also carry out parts of a workflow and use the results to continue.

This changes the relationship between knowledge and execution. Capabilities that once required separate research, implementation and troubleshooting can become accessible through a more integrated tool. For defenders, the important consequences are lower barriers to participation, faster iteration and cheaper repetition.

Knowledge becomes easier to apply

Public security information is abundant. Documentation, vulnerability reports, code examples and analysis are already widely available. Applying them to a particular system still requires effort: recognising which details matter, understanding dependencies and working through unexpected results.

An AI assistant can help bridge those gaps. An unfamiliar error message can be explained in context. A code fragment can be adapted to a different environment. A lengthy technical description can be reduced to the points relevant to the task.

For a less experienced operator, this can make unfamiliar work more approachable. For someone with established expertise, it can reduce the time spent on routine research and implementation. In both cases, the practical change is a shorter path from available information to an attempted action.

The UK’s National Cyber Security Centre identifies this broader shift in its assessment of AI’s impact on the cyber threat through 2027. It expects AI-enabled tools to extend access to intrusion capabilities and improve the effectiveness of existing techniques.

The feedback loop gets shorter

Technical work rarely follows a perfect sequence. A tool encounters an unexpected response, a dependency behaves differently or an assumption about the environment proves incorrect. Progress then depends on interpreting the feedback and deciding what to try next.

AI assistance can bring explanation, modification and another attempt closer together. The operator can ask about the observed result and receive a proposed adjustment without starting a separate search for each problem.

That matters because repeated adaptation is part of what makes technical work demanding. Reducing the effort involved in each iteration can make more attempts feasible, including attempts against systems that differ from the examples on which a tool was originally based.

For defenders, a failed attempt should therefore be treated as information about activity, rather than assurance that the operator has reached the limit of their capability. The next attempt may incorporate what the previous one revealed.

Assistance can become delegated execution

A conversational assistant returns information for a person to use. An agent connected to tools can go further: it can perform an action, inspect the output and continue through a sequence of tasks. The surrounding software determines which tools it can use, what information it receives and when a person must intervene.

This allows some of the coordination previously performed manually to move into the workflow itself. Several tasks may also run in parallel, with a person reviewing selected decisions and outcomes instead of executing every individual step.

In its report on an espionage campaign detected in September 2025, Anthropic describes an actor it assessed to be state-sponsored using an agentic system for substantial parts of the operation. Its account includes tool use, analysis and continued execution with intermittent human direction. It also records incorrect outputs, including invented credentials and mistaken claims about information obtained.

This is a vendor’s account of a particular campaign. It illustrates the operational significance of connecting models to tools: assistance with a task can develop into delegation of parts of the work. Actual results still need to be distinguished from the system’s own claims of success.

Lower costs change the economics of targeting

An attack has costs beyond the initial choice of a target. Someone must gather information, prepare or adapt tooling, deal with failures and decide whether further effort is worthwhile.

Automation can reduce the manual effort needed for some of those activities. The economic implication is that a wider range of targets and techniques may become worth attempting. An operation that would be unattractive if every step required sustained human attention may look different when parts can be delegated to software.

This does not require every attempt to succeed. Repetition can be useful to an attacker when the cost of unsuccessful attempts is low and occasional success has sufficient value.

The defensive implication is practical: an organisation’s exposure should be assessed through its reachable systems, credentials and potential consequences. Its perceived importance to a particular human attacker is an unreliable basis for deciding how much protection it needs.

The pressure extends beyond novel exploits

The most visible discussion often concerns discovering new vulnerabilities. Existing weaknesses also remain relevant: exposed services, known vulnerabilities, excessive permissions and poorly protected credentials.

NCSC’s assessment anticipates AI improving established intrusion methods and the exploitation of known vulnerabilities. Its forecast points to greater frequency and intensity of threats, rather than requiring every attack to introduce a new technique.

For application owners, that means familiar weaknesses can face more persistent and adaptable attention. Protection needs to hold across repeated attempts, different entry points and the operational changes made between deployments.

The previous article, Is the WAF Obsolete? Why Signatures Alone Are Not Enough, examines one part of that response: combining application traffic inspection with behaviour and context. The broader issue also concerns identity, infrastructure, software maintenance and the authority available after a component is compromised.

Make defensive work repeatable

Defenders benefit from accessible technical knowledge and automation as well. AI can help explain unfamiliar events, organise investigation material and prepare candidate checks for engineers to review. Its usefulness increases when the environment has reliable inventories, consistent telemetry and clear operational requirements.

Those foundations also make established defensive automation more effective. An exposed service should have an identifiable owner. A proposed configuration change should be checked before activation. A suspicious event should arrive with enough context for someone to investigate and act.

The same discipline applies when AI supports the work. A plausible explanation is a hypothesis to investigate. A proposed change needs validation. A report that an operation succeeded needs evidence from the affected system.

Reducing the repetitive work in those processes gives engineers more time to examine consequential decisions. The aim is a shorter, dependable path from observation to an appropriate response.

Plan for capability that is easier to obtain

AI is changing how technical knowledge can be acquired, applied and embedded in tools. Its security significance extends across the people attempting attacks, the workflows they use and the effort required to repeat an operation.

Organisations should plan for established techniques to become easier to attempt and quicker to adapt. Effective defence then depends on controls that remain useful under sustained automated activity: maintained systems, restricted access, proportionate traffic controls and an operational ability to recognise and contain abuse.

The technology will continue to change. The immediate engineering question is concrete: which parts of the defence depend on an attacker finding the work too difficult or too time-consuming, and how can those assumptions be replaced with enforceable controls?

References

• NCSC: Impact of AI on cyber threat from now to 2027
• Anthropic: Disrupting the first reported AI-orchestrated cyber espionage campaign

Related reading

• Is the WAF Obsolete? Why Signatures Alone Are Not Enough
• Protecting the WAF Itself