AI is accelerating attacks and changing cybersecurity and fraud prevention
Updated: 3 hours ago
Artificial intelligence is reducing the time needed to prepare cyberattacks, develop tools and build convincing conversations with victims. For businesses, this changes how quickly security teams need to respond. While a team is establishing what happened, an attacker may already be moving to the next stage. At ANTI-FRAUD FORUM 2026, Roman Sologub, CEO of ISSP, explored how organisations can reduce that delay and apply AI in cybersecurity and fraud prevention.
Minutes, not hours
CrowdStrike’s observations from 2025 illustrate the pace. Average eCrime breakout time was 29 minutes, with the fastest recorded case taking just 27 seconds. This measures the interval between initial access and an attacker moving to another system within the researchers’ dataset. It is not the duration of an entire attack, but it shows how quickly an intrusion can progress. CrowdStrike Global Threat Report 2026.
During that interval, defenders may still be working through routine tasks: finding the asset owner, comparing event logs, checking details with support and waiting for approval. Each step may be necessary, yet together they create a delay an attacker can exploit. Alongside detection speed, organisations should therefore measure the time to an action that actually reduces risk, such as checking a session, restricting access or stopping a harmful operation.
How AI changes the attacker’s work
AI helps attackers accelerate this sequence of activities. Experienced operators can understand unfamiliar environments, adapt code and test hypotheses more quickly. Someone lacking a particular technical or language skill can get assistance with tasks that previously required another specialist. Expertise still matters, but individual stages become less costly and more work can be carried out at once.
Adding tools and memory to a model makes an agentic workflow possible. The system retains context, assesses the outcome of one step and determines what to do next. Some tasks can run in parallel. For defenders, this means separate suspicious actions may belong to a coordinated process that advances faster than a team can investigate each event manually.
AI inside malware
Ukraine has already seen a language model used within malware. CERT-UA described LAMEHUG in an advisory about attacks on the security and defence sector. The malware used a model to generate commands from text descriptions, collect system information and locate documents for exfiltration. Part of its behaviour was generated during execution. CERT-UA advisory.
Google Threat Intelligence Group described other mechanisms. The experimental PROMPTFLUX used a model to rewrite and obfuscate code, while QUIETVAULT enlisted AI tools installed on a compromised system to search for additional secrets. These tools had different levels of maturity in the report and should not be grouped together as fully autonomous attacks. They do, however, show why a SOC needs visibility into software behaviour, tool use and data movement. Google Threat Intelligence Group research.
The risks of in-house AI assistants
The same concern extends to an organisation’s own AI assistants. Connecting an agent to email, documents or internal systems gives it access to information that may contain an attempt at manipulation. Through prompt injection, for example, an attacker can place instructions in a message or document to redirect the model’s behaviour. If the agent can take actions, the consequences can extend beyond an incorrect chat response.
An AI system’s capabilities therefore need to be accompanied by defined access rights, permitted operations and rules for handling data. Teams need to understand where a conclusion came from and what could happen if an input is malicious. MITRE ATLAS, a knowledge base of adversary tactics and techniques involving AI, can help organise and assess these scenarios.
Fraud and social engineering
Attackers can also manipulate the context in which a person makes a decision. In financial fraud, AI helps tailor messages to an individual, imitate a voice and sustain a convincing conversation. An employee may hear a familiar executive’s voice requesting an urgent transfer. A bank customer may receive a plausible explanation of why they need to change their access details or move funds immediately. Protection also depends on having a way to verify the request through an independent channel.
Ukrainian statistics show the cost of exploiting trust. According to the National Bank of Ukraine, the number of fraudulent payment card transactions resulting in losses fell by 5% in 2025, to 256,000. Yet losses rose by 24% to approximately UAH 1.4 billion, and social engineering accounted for 90% of the total losses. These figures cover payment card fraud overall and do not establish AI’s share. They demonstrate that fewer incidents can still cause greater harm. National Bank of Ukraine data.
Employee training should therefore help people act in specific situations. Short learning sessions, role-based scenarios and safe simulations provide opportunities to practise checking unusual requests and reporting concerns. Those reports must then reach the people who can connect them with other signs of an incident.
Shared context for SOC, fraud prevention and support
This is where cybersecurity and fraud prevention encounter a shared operational challenge. The SOC analyses devices, sessions and system events. The fraud team sees transactions, recipients and customer behaviour. Support hears the customer’s account of what happened. When this information remains in separate queues, each team works with only part of the story.
A new device can be entirely routine. But if it appears alongside changed access details, a transfer to an unfamiliar recipient and a report about a suspicious call, the risk assessment changes. Seeing that connection requires consistent identifiers, timestamps and access to the sources that explain the event.
In the model proposed by Sologub, teams and AI agents investigate different aspects of the same case using shared context. This does not necessarily require moving all data into one physical database. It requires access to relevant evidence within defined permissions, while preserving the ability to verify each conclusion.
What AI agents can take on
With that foundation, agents can take on part of the investigation. Rules-based automation works well when the route is known in advance. An agentic approach is useful when the next step depends on what has already been established. An agent can reconstruct a timeline, identify a contradiction, consult another source or refer an ambiguous case to a person.
Microsoft Perception, discussed in the presentation as Project Perception, illustrates this direction. The system brings together specialised agents to expose weaknesses, investigate threats and strengthen security. Their findings become shared context, while the team retains control over critical decisions and can stop execution. Microsoft documentation.
The organisation defines the limits of that authority. Being able to gather evidence does not automatically grant permission to block a payment or an account. Each action needs clear conditions and an approval path. Performance should be assessed against both speed and quality: an incorrect block imposed in a second can create a new business problem.
ISSP case: fraud-report analysis
ISSP’s own work provides a practical example of using shared context. At the forum, Roman Sologub presented a fraud-report analysis pipeline developed by the company. It processes call transcripts and support agents’ notes. After records are de-identified, the system classifies reports, supports its conclusions with excerpts from conversations and identifies the reported fraud scenario. Related reports can be grouped into a single incident.
Separate checks assess transcript quality and the support agent’s work. If the agent’s classification differs from the AI analysis, the discrepancy receives an additional review. Analysts receive a structured case containing evidence and information gaps. This provides a basis for examining changes in fraud tactics and improving the way reports are handled.
Where to start
A process like this offers a concrete starting point for AI adoption. An organisation can select a workflow with substantial manual effort or waiting time, establish its current performance and define quality criteria. A pilot can then show whether the team reaches a justified action faster and how that affects errors and workload. These verified improvements help build a defence that can keep pace with accelerating attacks.
