How CISOs can prepare for the next generation of AI-driven security

It's been a rough patch for software security. The first couple of months of this year brought several high-profile supply chain incidents, the kind that remind everyone how fragile the software supply chain really is. Team PCP, the attackers responsible for hitting Trivy, LiteLLM, and countless other projects at the beginning of this year, labelled 2026 'the year of the supply chain.’ Then, Anthropic shocked the world with its new security-focused Mythos model, which is so powerful that it was held back from general release. At the same time, the company introduced Project Glasswing, an effort meant to give organisations running critical infrastructure a head start on finding and patching before Mythos became generally available and other advanced models followed. 

Since the model's debut, reactions have been mixed. Mythos has often been referred to as an “inflection point,” and the moment is described as a global arms race for AI security. Others have framed it in terms of major geopolitical fallout.

Among its first 50 users, Mythos uncovered more than 10,000 critical- or high-severity vulnerabilities in just one month. These are vulnerabilities that had previously remained undetected by software scanners. Yet, the more powerful aspect of the Mythos model compared to others isn't necessarily identifying security holes; it's how it can exploit those holes and combine them to have an even greater effect. This creates a double-edged sword for the defender, because attackers have the same opportunity to use these models to discover vulnerabilities and exploit them more quickly than ever before. 

Ultimately, attackers and defenders have access to the same models, and there's a reasonable argument that defenders can protect their systems before attackers exploit them if they use the tools wisely. This moment has been a call for the industry to rethink its software stacks and vulnerability management systems from the ground up. 

So, what should engineering teams do?

If your organization has access, you can start by using the models yourselves. Regardless of whether your codebase is proprietary or open source, have agents review pull requests from a security perspective and run them against legacy code to flag potential weaknesses.

Don’t change the fundamentals. Defense in depth still matters. So does least privilege. Build in multiple layers, isolate where you can, and get rid of long-lived access tokens wherever possible, which were a factor in both the Trivy and Axios incidents.

Open source carries a particular risk here, since the code is sitting in the open for anyone, including a model, to study. And open source underpins nearly every supply chain. Teams need the latest patches fast, but they also need to be sure those patches are coming from a trustworthy source, not a compromised update dressed up to look legitimate.

To secure themselves, organisations must improve the integrity of their software supply chain by using secure-by-design methodologies during development, enabling them to verify and trust all parts of the software. Software teams need to trust their sources. That means knowing exactly where open source components come from, building and deploying through secure pipelines, and monitoring dependencies throughout the software's lifecycle. The tools already exist: reproducible builds, automated vulnerability tracking, authenticated patches, and faster patch cycles. Used together, they keep compromised code out of production.

Several industry collaborations are also focused on strengthening open source security against AI-powered threats. Athena is one such coalition for the orchestrated defense of open source software, in which its members feed vulnerability findings from across the industry into Athena, which then takes each one through its full lifecycle and turns it into protection for everyone.  

The question is no longer whether AI will transform cybersecurity - it already has. Vulnerabilities that once took weeks or months to discover and exploit can now be identified in hours. Organisations that continue relying solely on reactive scanning and manual patching will struggle to keep pace.

The winners will be those that build trust into every stage of the software supply chain - from the source code they consume to the software they deploy. Preparing for advanced AI isn't simply about adopting new security tools; it's about establishing a software foundation that both developers and AI Agents can verify and trust at machine speed. Those that invest in trusted software supply chains today will be best positioned to innovate with confidence, accelerate AI adoption, and compete in the AI-first era.
 

Naveen Sharma

---

Naveen Sharma is Global Vice President - Partnerships at Chainguard