In AI security, there’s no room for a defender’s mindset

In June, the Five Eyes intelligence alliance issued a joint statement warning that frontier AI models are expected to surpass current industry expectations, fundamentally altering both offensive and defensive cyber capabilities. This isn't theoretical; evidence shows that attackers can already bypass safeguards. The skill level required for unsophisticated attackers to execute increasingly sophisticated attacks continues to decline.

In this rapidly evolving environment, defending attack surfaces is no longer sufficient. Instead, security professionals must adopt an offensive mindset. The Five Eyes' warning signals that existing security approaches are outpaced by adversaries' exponential gains in speed and scale. In this context, the best defense is a robust offense.

Security professionals often instinctively view powerful model releases as detrimental news. However, this mindset fails to address the evolving security landscape. Most security defenses were originally tools designed for exploitation, which security teams adapted for defense purposes. The same logic applies to AI-powered security, and defenders possess advanced technologies akin to their attackers.

Engineering backgrounds, particularly in software companies, tend to instill an attacker mindset, characterized by proactive hole hunting across entire programs, mirroring adversary tactics. This is the mindset needed in the current AI security landscape.

To cultivate an attacker mindset, security teams should focus on building scalable solutions that automate outcomes for the business while continuously identifying potential shortcomings in the program, threat landscape, and infrastructure and asset catalog. This requires genuine curiosity about how systems are set up, why they function that way, and what could go wrong.

Security research thrives on this curiosity, which involves understanding system configurations, anticipated behavior, and ways to achieve unintended outcomes. Engineering-led teams can adopt automated agents that identify, triage, remediate, and report on risks, aligning with specific business outcomes.

To maximize the AI advantage, two practical steps are essential. First, maintain model neutrality by constructing workflows unbound by specific models or vendors. The leading models and AI tools may become obsolete within six months, so model neutrality ensures the ability to deploy in air-gapped or self-hosted environments when necessary, rather than merely swapping vendors.

Second, limit the scope of agentic tasks to well-defined problems. Providing an agent with a broad, vague task leads to short-term success followed by a loss of focus and ineffective results. Conversely, assigning specific tasks with clear expectations ensures consistent, intelligent outcomes. By adopting this approach, security programs can transition beyond reactive measures and proactively meet the demands of an AI-driven cyber landscape.