Cybersecurity 2026: autonomous malware and deepfakes, how to protect your company

Artificial Intelligence is changing business cybersecurity from both sides. Defensive teams can use it to analyze large volumes of information, identify anomalies, and accelerate incident response. Attackers have the same capabilities to research targets, customize social engineering campaigns, develop code, and produce fake digital identities more quickly.

The Global Cybersecurity Outlook 2026 from the World Economic Forum precisely identifies AI as the main factor transforming cybersecurity during 2026: 94% of respondents expect it to be the force with the greatest impact on the sector. Additionally, 87% identified AI-related vulnerabilities as the cyber risk that grew the fastest during 2025.

For companies, this evolution modifies a significant part of the defensive model. Recognizing a poorly written email, detecting a strange voice during a call, or trusting that known malware will be identified by a signature is becoming increasingly insufficient.

Attacks can better adapt to the victim's context and produce more convincing synthetic communications. The business response needs to combine technology, identity, verification processes, and reaction capability.

How artificial intelligence is changing cyberattacks

The use of AI by malicious actors can already be observed at different stages of an attack. Microsoft Threat Intelligence documents its use for reconnaissance, phishing generation, creating fake identities, analyzing stolen information, and developing or debugging malware.

The main change is related to speed and scale. Tasks that previously required more time or specific knowledge can be accelerated through generative models.

This reduces some operational barriers for attackers and allows relatively traditional campaigns to incorporate higher levels of personalization.

Social engineering can be much more convincing

Grammatical errors, poor translations, and generic messages have served for years as signals to detect phishing attempts. Generative models can eliminate many of those signals.

An attacker can use public information about a person to generate a message related to their position, company, colleagues, or recent activity. They can also adapt the language and communication style to the victim's context.

Microsoft has precisely observed the use of collected data about targets to personalize messages, as well as the generation of documentation, identities, and fake professional communications.

The business risk no longer solely depends on mass attacks. The same technology that allows generating thousands of messages also facilitates building much more specific communication for a person with privileged access.

Autonomous malware: what really exists in 2026

The concept of autonomous malware describes a potential evolution where malicious software can use Artificial Intelligence during its execution to adapt certain actions to the compromised environment.

It is important to differentiate this capability from malware developed with AI assistance.

Currently, Microsoft indicates that much of the malicious use observed consists of using models to produce code, debug malware, generate scripts, or accelerate engineering tasks. The attacker continues to control the targets, victim selection, and fundamental decisions of the operation.

From AI-assisted malware to adaptive malware

The next stage begins when the model is no longer used solely during development and becomes part of the malware's operation.

Microsoft collects public research on experimental families capable of invoking models during execution to dynamically generate scripts, modify code, or obfuscate certain operations. These techniques still present limitations related to reliability, latency, and operational risk.

Their importance lies in the direction they indicate.

Malware capable of analyzing its environment and modifying certain parts of its behavior can reduce the effectiveness of defenses that are overly dependent on static indicators. This reinforces the need to detect anomalous behaviors in addition to known files or signatures.

From signatures to behavior analysis

Signatures remain useful for identifying known threats, but a business strategy needs to incorporate controls capable of observing what is happening within endpoints, identities, and networks.

A process that begins to execute unexpected commands, access credentials, establish unusual connections, or modify large amounts of information can generate signals even if the file used does not yet match a known signature.

The Cybersecurity Specialists can combine protection technologies with risk analysis, hardening, monitoring, and response procedures tailored to each organization's infrastructure.

This approach is especially relevant when offensive tools can generate variants or modify components more quickly.

Deepfakes turn identity into an attack surface

Deepfakes shift the impact of AI from malware to human trust. Synthetic audio, video, and images allow for the construction of scenarios where a person appears to be seeing or hearing someone familiar.

The World Economic Forum identifies audio and video deepfakes, along with forged documentation, as part of the evolution of fraud and social engineering driven by AI.

Microsoft has also documented the use of voice cloning, real-time voice modification, and video overlays to impersonate identities or create fake personas.

In a business context, the target may be an employee in finance, human resources, IT, management, or anyone with the ability to approve a sensitive action.

A convincing video should not serve as proof of identity

Defense against deepfakes requires modifying some business procedures. An organization cannot assume that hearing a manager's voice or seeing them via video call constitutes sufficient validation to authorize a critical operation.

The applicable principle is to verify the request through an independent channel.

If a person requests an extraordinary transfer via a call, the procedure may require confirming the operation using a previously registered corporate number. If a vendor requests a change to their bank account, the company can verify the change using the contact details existing in its systems and not those provided in the same message.

Technology can help detect synthetic content, but processes should remain secure even when the deepfake manages to bypass that detection.

How to act in the event of a potential deepfake impersonation

Companies should establish known procedures before an incident occurs. When a request involves money, credentials, sensitive data, or significant changes, certain signals warrant additional verification:

• Requests for transfers or bank changes outside the usual procedure;

• Urgent requests to share passwords, MFA codes, or confidential information;

• Instructions to avoid contacting other people in the organization;

• Sudden changes in channels, phones, or accounts typically used;

• Requests to install software or provide remote access;

• Pressure to complete an operation before being able to verify it through another channel.

No single signal proves the existence of a deepfake. Their function is to trigger a verification procedure that reduces reliance on appearance, voice, or message content.

Phishing and business fraud benefit from the same technology

Deepfakes and malware represent only a part of the landscape. AI is also increasing the ability to build more convincing phishing and fraud campaigns.

The World Economic Forum notes that digitally enabled fraud has become the primary cyber concern for surveyed CEOs for 2026, while ransomware remains the main concern among CISOs.

This difference is significant. For security, a ransomware infection can halt critical systems. For management, a fraud campaign can lead to significant financial losses without needing to technically compromise the entire infrastructure.

Business cybersecurity needs to cover both dimensions.

Business Email Compromise can incorporate new forms of impersonation

Business Email Compromise traditionally uses compromised accounts or addresses similar to legitimate ones to request payments, bank changes, or information.

AI can enhance these campaigns by generating messages tailored to the style and context of an organization. Synthetic voice or manipulated video can add a second layer of credibility when the attacker needs to overcome doubts.

For this reason, it is risky to simply use a call as a method of confirming a suspicious email. If both channels can be controlled or simulated by the attacker, the company needs a previously established independent mechanism.

What AI-related threats should a company watch for

The capabilities observed in 2026 affect different parts of corporate security. Not all require completely new tools, but they do require reviewing how existing controls function.

The table shows a common pattern: identity is becoming one of the most important defensive layers. If an attacker manages to impersonate a person or use their credentials, they can bypass numerous technical barriers.

Protecting identity must be a priority

Strong passwords remain necessary, but they are insufficient when a credential can be obtained through phishing, malware, or leaks.

Multi-factor authentication adds an additional layer, although there are also attacks designed to capture sessions or get the victim to approve a request.

Organizations should move towards phishing-resistant mechanisms when possible and pay special attention to administrative accounts and users with access to sensitive information.

The teams responsible for IT Infrastructure need to coordinate identity, devices, networks, and applications to reduce the possibility of a compromised credential providing excessive access.

Least privilege limits the impact of a compromised identity

The principle of least privilege states that a person, application, or service should only have the permissions necessary to perform its function.

This practice becomes more important when attackers can accelerate reconnaissance and analysis after gaining initial access.

A compromised account with limited permissions offers fewer opportunities to move through the infrastructure than an identity with broad access to systems that it does not need to use regularly.

Periodic review of privileges also allows for the removal of accumulated accesses when a person changes roles or no longer needs certain resources.

EDR and XDR help observe anomalous behaviors

The evolution towards more dynamic malware increases the value of technologies capable of analyzing behavior within endpoints and correlating signals from different systems.

EDR can log processes, connections, modifications, and other device events. XDR extends correlation to other layers such as identity, email, Cloud, or network depending on the platform used.

These tools do not replace a secure architecture. Their function is to enhance the ability to identify suspicious activity and provide context for investigating a potential incident.

The quality of detection also depends on configuration, available telemetry, and the team's ability to respond to generated alerts.

The speed of attack increases the importance of patching

AI can help attackers investigate systems and adapt tools more quickly. This reduces the available margin between the disclosure of a vulnerability and its potential exploitation.

Organizations need to know what assets they have exposed, what software they use, and what vulnerabilities actually affect their systems.

A vulnerability management program should prioritize based on risk rather than merely accumulating lists of vulnerabilities. Internet exposure, asset criticality, availability of exploitation, and associated privileges can significantly alter priority.

Automation can help identify and deploy updates, but certain critical systems require testing procedures to reduce operational risk.

Business AI also creates a new attack surface

The risk associated with Artificial Intelligence does not only come from attackers using models. Companies themselves are incorporating assistants, RAG, copilots, and agents into their operations.

Each new system can introduce connections with data, applications, APIs, and corporate credentials.

The World Economic Forum notes that the proportion of organizations evaluating the security of their AI tools increased from 37% in 2025 to 64% in 2026. The evolution shows that the security of these systems is becoming part of business practices.

Teams developing Artificial Intelligence solutions need to consider security from the architecture stage and not just after deploying the application.

AI agents need specially controlled permissions

An agent may have the ability to consult documents, use APIs, modify records, or execute other tools. If their credentials provide overly broad permissions, a failure or compromise can have greater consequences.

Therefore, agents should operate through specific identities, minimal permissions, and clearly defined tools.

It is also advisable to log what actions they perform, what systems they use, and what decisions require human intervention. The ability to automate tasks should grow alongside the capacity to control that automation.

What controls should a company prioritize in 2026

The emergence of offensive AI does not mean that all previous defenses have ceased to function. A considerable part of protection still depends on correctly executing fundamental controls and adapting them to the new scenario.

A business strategy should review at least:

• Identity and access: phishing-resistant MFA, least privilege, and enhanced protection for administrative accounts.

• Endpoints and servers: EDR/XDR, hardening, updating, and reducing unnecessary software.

• Vulnerabilities: asset inventory, risk-based prioritization, and defined patching cycles.

• Fraud and deepfakes: independent verification for payments, bank changes, credentials, and sensitive operations.

• Backups: isolated, protected copies that are periodically tested for restoration.

• Corporate AI: inventory of tools, permissions, data protection, and risk assessment before integrating with critical systems.

• Incident response: responsibilities, procedures, and alternative channels prepared before an attack occurs.

Technology covers only part of this strategy. Clear processes and known responsibilities reduce the time needed to make decisions during an incident.

Employee training also needs to evolve

Traditional awareness programs often teach how to detect spelling errors, suspicious domains, or unexpected messages. These signals remain useful, but AI-generated attacks can be much more convincing.

Training should progressively focus on procedures.

An employee does not need to become an expert in detecting deepfakes. They need to know that certain operations should never be approved solely via a call or email, regardless of how convincing the person seems.

They should also know whom to report to when they suspect a request and have a simple mechanism to do so.

A security culture works best when reporting a potential threat is quick and does not penalize the employee for having doubts.

Preparing for the incident is as important as trying to prevent it

No organization can guarantee that it will block all attacks. Therefore, business cybersecurity needs to combine prevention with resilience.

An incident can compromise an account, device, or application without it having to automatically lead to widespread disruption.

Segmentation, least privilege, backups, monitoring, and response procedures help limit the impact and recover operations.

It is also advisable to conduct periodic exercises. Simulating a compromised administrative account, ransomware, or a fraudulent request via deepfake allows checking if responsibilities and communication mechanisms actually work.

Cybersecurity 2026 demands verifying more and trusting less in appearances

Artificial Intelligence is providing attackers with tools to work with greater speed, scale, and customization capability. More convincing phishing, synthetic identities, deepfakes, and AI-assisted malware development are already part of the threat landscape.

Completely autonomous malware still needs to be treated with precision. The available evidence mainly shows human attackers using AI as a capability multiplier and initial experiments with malware capable of incorporating models during execution.

For companies, waiting for that autonomy to reach greater levels would be an unnecessarily reactive strategy. Many of the necessary defenses can be implemented now.

Identity, least privilege, endpoint protection, vulnerability management, independent verification, security of AI applications themselves, and incident preparedness form a much more resilient defense than relying on employees or tools to always recognize what content is authentic.

The evolution of threats is also increasing the need for specialized talent. When an organization does not have all the necessary capabilities internally, an Outsourcing IT model can allow for the incorporation of cybersecurity, infrastructure, and Artificial Intelligence specialists to strengthen specific areas without immediately building a complete team.

In 2026, protecting a company means assuming that an email can be perfectly written, a voice can sound authentic, and a video can show a person who was never in front of the camera. The response is to design systems and processes where a convincing imitation is not enough to gain access, money, or information.