Laurent Giraid is a seasoned technologist whose career has been defined by a deep exploration of the ethical and practical applications of machine learning. With a background that spans natural language processing and high-stakes infrastructure, he has become a leading voice on how frontier AI models can be harnessed to protect global systems. In this conversation, we delve into his analysis of the recent security shifts at the highest levels of the financial sector, focusing on how large-scale networks are moving away from traditional defense toward agentic, AI-driven security architectures.
The discussion centers on the evolution of cybersecurity from static pattern matching to reasoning-based offensive and defensive modeling. We explore the transition from legacy metrics like simple vulnerability counts to more dynamic measurements of resilience, such as adaptation time and the closure of complex exploit chains. Giraid also provides insights into the emerging risks of agent-led commerce, the necessity of securing the open-source supply chain, and why the current window for organizations to outpace machine-speed attackers is rapidly closing.
When using frontier AI to identify exploit chains rather than isolated bugs, how do you manage the trade-offs between depth of analysis and the speed required for remediation?
The shift from finding isolated bugs to identifying full exploit chains is a monumental change in how we perceive risk. When you use a model like Anthropic’s Claude Mythos to examine a network that connects nearly 5 billion payment credentials to over 175 million merchant locations, the sheer volume of data can be overwhelming. To manage this, the analysis cannot be a simple scan; it must be a governed pipeline that prioritizes attack surfaces through active threat modeling before the deep-dive analysis begins. By focusing on how minor weaknesses stitch together across a stack that moves money in roughly 160 currencies, we can move faster because we are no longer chasing thousands of low-impact alerts. Instead, we use a multi-agent deterministic voting system to ensure that only the most credible and dangerous chains advance to the remediation stage, which compresses the entire lifecycle from discovery to a validated fix that developers can actually deploy.
How does the philosophy of “pessimism and paranoia” reshape the way a massive financial network approaches its defense-in-depth strategy in the age of AI?
Pessimism and paranoia at this scale mean assuming that your hardened defenses will eventually fail and designing for that exact moment. For a network operating across more than 200 countries and territories, traditional layered defenses and zero-trust architectures are just the baseline. When frontier models were tested under Project Glasswing, they identified more than 10,000 high- or critical-severity vulnerabilities in just the first month across various critical systems. This reality forces a rethink of our assumptions because it shows that even decades of hardening can be bypassed by AI-speed reasoning. We have to move toward an “agentic defense” where the system is constantly trying to break itself, identifying the logic and data flow gaps that exist between traditional security signatures that legacy tools often miss.
Why is the move toward an open-source “agentic harness” so significant for the broader security community, and how does its structure differ from traditional security tools?
The release of the Visa Vulnerability Agentic Harness on GitHub, which already has 595 stars and 97 forks, represents a shift toward transparency in critical infrastructure security. Unlike a monolithic scanner that just gives you a list of problems, this harness is a pipeline consisting of four phases and eleven stages, ranging from code ingestion to fix validation. It is designed to be multi-model, which is crucial given that 59% of enterprises plan to adopt or switch their agent security tooling within the next year. By providing a reference implementation that combines deterministic controls with AI-informed reasoning, it allows any team to inspect and adapt the architecture to their specific needs. This structure ensures that human oversight is baked into every stage, preventing the AI from making autonomous changes without a verified reasoning chain.
Can you elaborate on the transition from legacy security metrics like CVE counts to the more outcome-oriented “Mean Time to Adapt” (MTTA)?
Legacy measures like mean time to detect or raw CVE closure counts are increasingly becoming “vanity metrics” that can look good on paper while actual exposure grows. The data shows that fewer than 1% of CVEs are ever actively exploited, yet teams spend thousands of hours patching bugs that may never be part of a viable attack path. Mean Time to Adapt (MTTA) shifts the focus to three dimensions: inventory freshness, the number of exploitable paths per release, and the validation cycle time for a fix. It forces a security team to prove that an end-to-end attack chain is actually closed, rather than just showing that a patch was applied to a single component. This approach assumes that every exploitable path will eventually be exercised in production, making it a much more honest assessment of a company’s true security posture.
With the rise of agentic commerce, what are the specific identity and security challenges when AI agents begin transacting on behalf of humans?
We are entering a future where AI agents will not just assist us but will actually complete financial transactions, which introduces a massive identity crisis. Current research indicates that 69% of enterprises already have credential sharing occurring somewhere in their agent deployments, which is a significant vulnerability. Companies that use shared credentials report security incidents at a rate of 63.5%, compared to only 40.9% for those that use individual, scoped identities for each agent. To solve this, we have to treat AI agents as distinct identities within our governance frameworks, requiring full audit trails and least-privilege enforcement. We are building the trust frameworks and “agent readiness scoring” today because if we don’t, the fraud scenarios we see in simulation environments will become a daily reality for millions of consumers.
How is the risk of the global technology supply chain being mitigated through initiatives like Project Lightwell?
No matter how strong your internal defenses are, you remain exposed through every weak vendor or open-source component in your stack. Project Lightwell is a $5 billion initiative involving major players like IBM, Red Hat, and several global financial institutions to harden these shared components through AI-driven validation. We are moving toward a world where AI-specific security posture is a non-negotiable part of supplier due diligence, requiring continuous vulnerability validation and living software bills of materials. The goal is to ensure that the MTTA clock doesn’t stop at our own perimeter; we need our entire ecosystem to adapt at the same speed. This collective defense is the only way to secure the underlying software that supports billions of daily transactions worldwide.
What are the “non-negotiable” architectural practices that you believe every security team should be prioritizing right now?
The white paper detailing the harness architecture lists 12 non-negotiable practices, but three stand out as the heaviest lifts for any organization. First, you must shift security so far left that exploitable flaws are designed out before the code even reaches production. Second, there must be a ruthless commitment to replacing high-risk or under-supported components before they become material exposures. Finally, and perhaps most importantly, defenses must be refactored to run autonomously under human governance so that response times can keep pace with machine-speed attacks. This isn’t just a suggestion for those with massive budgets; it’s a blueprint for any team to start using structured triage artifacts to compress their discovery-to-remediation lifecycle today.
What is your forecast for the evolution of agentic security over the next five years?
I believe we are rapidly approaching a “security singularity” where the volume and sophistication of machine-led attacks will completely outstrip human capacity to respond in real-time. Over the next five years, the “Mean Time to Adapt” will become the singular metric that determines an organization’s survival, and we will see a total migration away from provider-native controls, which 82% of enterprises currently rely on. We will see the birth of fully autonomous security agents that not only detect and fix vulnerabilities but also negotiate security protocols with other agents across the supply chain. The window to build these governed, agentic defenses is open right now, but it will close as soon as the models behind the attacks reach the next level of reasoning. Organizations that fail to implement these architectural shifts today will find themselves defending a 21st-century network with 20th-century tools.
