Why cybersecurity leaders see hope in AI-powered hacking

Advanced AI models offer proactive cybersecurity benefits

AI is already reshaping cybersecurity, and its next frontier is automation. Anthropic’s Claude Mythos Preview is a prime example of how artificial intelligence is evolving from a support tool into an active partner in defense. The model shows strong capability across areas that matter most for large organizations, mathematics, code analysis, security engineering, and automated vulnerability detection. It can autonomously discover and simulate potential exploits in ways human teams could not achieve at scale or speed.

Alexander (Sacha) Babuta, Director at the Centre for Emerging Technology and Security (CETaS) at the Alan Turing Institute, points out that this same capability could become a critical advantage for enterprises. Rather than fearing autonomous systems that find vulnerabilities, companies can deploy them internally to perform rapid system self-checks and apply instant fixes. This flips the dynamic of cybersecurity from reactive defense to proactive protection. Using AI this way ensures that the same tools that might one day threaten security become essential to strengthening it.

For companies handling complex digital infrastructure, particularly those operating across multiple systems, countries, or compliance environments, this kind of AI-driven approach creates measurable advantages. Faster incident detection reduces downtime. Automated patching cuts response time. Ultimately, it redefines cybersecurity workflows as leaner, smarter, and more adaptive.

Decision-makers should understand that this technology is operational now, and early adoption provides both technological and strategic benefits. Businesses prepared to integrate AI-based vulnerability detection will gain the upper hand in resilience and cost efficiency, keeping their digital systems a step ahead of emerging threats.

“Dark AI” phenomena in cybercrime remain marginal despite early enthusiasm

There is growing discussion about the rise of so-called “dark AI”, custom or jailbroken large language models (LLMs) developed for use in cybercrime. These models are promoted within underground forums as powerful hacking tools. However, real-world evidence from 2022 through 2025 shows very limited impact. Most of these systems exist as small, homegrown experiments rather than credible threats at scale.

Ben Collier, Senior Lecturer at the University of Edinburgh, has led research examining these developments in cybercrime communities. His findings reveal that while online discussions about AI-assisted hacking have increased, few participants have the technical capability to use such tools effectively. Many users attempt to apply legitimate AI systems such as ChatGPT and Claude to automate basic processes, but they tend to struggle with complex or high-risk operations. Collier describes most of the activity as basic logistics work, administrative coding, data management, or scheduling, similar to what any small team might perform when managing routine tasks.

This observation is important for business leaders. It means that the current generation of AI tools has not empowered a new wave of highly capable cybercriminals. The threat remains, but its scale and sophistication are overstated. Most users experimenting with these models are hobbyists or low-skill operators.

Executives should still take these developments seriously but keep perspective. While “dark AI” exists, its current influence on actual cyberattacks is minimal. The real challenge lies in tracking how these capabilities evolve and preparing corporate infrastructure to stay secure as tools become easier to use and more technically advanced. A measured understanding of both the current limitations and future risks will allow organizations to act strategically rather than react impulsively.

Evolving AI capabilities may enable complex autonomous cyberattacks

The next phase of artificial intelligence brings power and uncertainty in equal measure. Models capable of writing code and troubleshooting systems can now perform entire cyber operations autonomously. Recent work from the AI Security Institute (ASI) demonstrated this clearly. A frontier AI model successfully completed a 32-step cyberattack on a simulated corporate network, moving from reconnaissance to full network takeover without human intervention. Adam Beaumont, Interim Director at ASI and former Chief AI Officer at GCHQ, explained that this task would normally demand around 20 hours of skilled human work. The model achieved it far faster and without explicit instructions for each phase.

Beaumont emphasized that this demonstration was not a test of the model’s response behavior, it was direct execution. The system took autonomous actions to achieve a defined objective, an early indication that advanced AI can operate beyond conventional task boundaries. That raises fundamental questions about control, intent, and oversight. As Beaumont noted, current understanding of how to ensure these systems remain under meaningful human control is still incomplete. He described the ASI demonstration as an “honest starting point” for evidence-based governance.

For executives, this shift represents both opportunity and risk. On one hand, automation at this level could redefine cybersecurity efficiency. On the other, it highlights how quickly AI can surpass human pacing and capability in sensitive contexts. Governments, regulators, and enterprises will need to align on safety protocols and establish verification systems that validate AI behavior before deployment.

Strategic investment in AI literacy, policy design, and interdisciplinary collaboration will be essential. Business leaders who understand the mechanics of this technology will be better positioned to balance innovation with control. The goal is not to limit progress but to guide it responsibly, building clarity, safety, and accountability into AI development from the outset.

Key takeaways for leaders

AI-driven cybersecurity as a proactive defense: AI models like Anthropic’s Claude Mythos Preview can autonomously detect and address system vulnerabilities before attackers exploit them. Leaders should invest in AI-assisted security tools to strengthen digital resilience and reduce response time.
Limited real-world impact of “dark AI”: Research shows that cybercriminal use of jailbroken AI models remains minimal due to low technical proficiency. Executives should focus resources on legitimate AI adoption rather than overreacting to unproven external threats.
Growing need for oversight in autonomous AI systems: Recent demonstrations show that advanced AI can execute full-scale cyberattacks without human input, highlighting both risk and potential. Leaders should push for governance frameworks and internal AI control measures to maintain security and accountability as autonomy increases.