The arrival of agentic AI models in 2025 will mark the transformation of artificial intelligence from helpful assistants into peers, or even experts, that can solve problems unaided, and force security teams to rethink defence strategies, according to new research from Malwarebytes.

According to the cybersecurity provider’s 2025 State of Malware report, the introduction of AI agents – AI models like OpenAI’s “Operator” that can plan, act, reason, and use tools – could finally deliver the cybersecurity disruption that many expected after ChatGPT’s release in 2022.

Arguing that AI agents could come this year, the research points to public statements made by the biggest AI firms like Anthropic, OpenAI and Google DeepMind, which claim that the development of artificial general intelligence – the step beyond agentic AI which could match or exceed human intelligence – is only between two and ten years away from a reality.

The study highlights the so far limited impact of AI on the malware landscape, due in large part to the current capabilities of available models. 

For example, a report from OpenAI published in October found that despite persistent attempts to use ChatGPT to perform malicious activities – such as writing and debugging malware, or setting up command and control infrastructure – the AI model offered threat actors “limited, incremental capabilities for malicious cybersecurity tasks”. 

Although there is ample evidence that AI is helping threat actors to refine their efficiency,  Malwarebyte’s study points out this hasn’t translated into the emergence of brand new capabilities, but that could all change with agentic AI.

The Threatdown report predicts that AI agents will not only narrow the skills gap for threat actors, but could be used to scale up the number and speed of attacks.

According to the study, that could mean ransomware gangs using agentic AI to attack multiple targets at one time using evolving code, malicious AI agents being tasked to search out and compromise vulnerable targets, or even AI capable of running and fine-tuning malvertising campaigns in real time.

However, the report notes that these AI agents could also become a critical defence tool for cybersecurity teams.

Malwarebytes said that as AI becomes more capable, security teams could begin delegating more of their tasks to autonomous agents with minimal instructions, allowing networks to keep up with constantly evolving threat tactics.

That could mean AI agents tasked with looking for supply-chain vulnerabilities, keeping a running inventory of internet-facing systems to ensure they’re patched, or monitoring a network overnight to respond immediately to suspicious alerts.

“To combat threat actors using AI to make attacks more potent, organizations need to operate their security as efficiently as possible,” recommends the report.

“Security teams will have to embrace AI-powered threat detection and defence strategies to keep up with the evolving, agent-enhanced threat landscape.”

Recommended reading

• Capgemini: How to Build a Responsible Generative AI Strategy
• MIT Tech Review: Responsible AI Principles Urgently Needed
• Report: 58% of Business Leaders Have No AI Training

The rise of agentic AI is increasingly being recognised as a potential problem, and not only in cybersecurity circles.

Research from Pegasystems shows that a third of workers in the UK and the US do not trust AI agents’ reliability or quality, with similar numbers noting a lack of emotional intelligence and human intuition from AI agents, as well as three in ten reporting that they do not trust the accuracy of AI agent’s responses.

Despite these concerns, most IT executives see value in scaling the use of agentic AI in their firms, with UiPath’s 2025 Agentic AI Report finding that 90% of IT executives believe business processes would be improved though its use, and over three-quarters (77%) planning to invest more developing agentic AI.