By Chaim Mazal, GitLab CISO
Advanced models and agentic systems are compressing the time and cost required to find and exploit weaknesses. The risks themselves are familiar: an open server, an over-scoped credential, or an exposed deployment path.
To keep pace, organisations need AI resilient security programmes that scale with how this technology evolves. AI resilience is just cyber resilience moving towards an attacker-first, engineering-led model.
The underlying work is not new, but the mindset is. A defender mindset waits for identified activity to inform decisions. An attacker mindset proactively searches for gaps across the programme, the threat landscape, and the asset estate.
That is why I encourage my teams to think proactively by looking at AI through an adversarial lens and finding ways AI can strengthen defences. Nearly all security defences started as tools built to exploit environments, and security teams later repurposed them for defence. As cybersecurity professionals, we have more context than the attacker, greater access to our own systems, and a chance to make the construction of software itself observable and governable. We have to use that advantage.
To do so, teams should first focus on model neutrality, building workflows that are not tied to one model or vendor, since the best models and AI tools today may not be right in six months. Second, the scope for agentic tasks needs to be narrow and well defined. Give an agent a large, vague problem, and it performs well briefly, then loses the plot and starts generating whatever it thinks will satisfy the prompt. Give it a specific task and prioritise delivering the right context up front for it to succeed consistently.
Beyond that, the fundamentals still matter: identity and access management, secure software development, governance, monitoring, and the ability to recover systems and data. But those fundamentals cannot operate as disconnected pillars in an AI-driven landscape.
The biggest obstacle – and risk – organisations face in moving forward right now is analysis paralysis. To start, assess each area of focus, define the deterministic outcome you want to achieve, and work backwards using the technology already available.