What Happened
In an unprecedented demonstration at a cybersecurity conference in March 2026, Nicolas Carlini, a Research Scientist at Anthropic, showed Claude AI discovering zero-day vulnerabilities in real-time. The AI successfully identified:
- A blind SQL injection vulnerability in Ghost CMS (CVE-2026-26980) that allowed complete admin database compromise
- A complex stack buffer overflow in the Linux kernel’s NFSv4 daemon that had existed undetected since 2003
- Multiple smart contract vulnerabilities worth millions in simulated funds
Carlini, who has published extensively on AI safety and adversarial machine learning, admitted during the presentation that Claude’s vulnerability discovery capabilities now exceed those of expert human researchers.
The Ghost vulnerability was particularly striking because the popular publishing platform, with 50,000 GitHub stars, had never experienced a critical security vulnerability in its entire history. Claude found and exploited the flaw in just 90 minutes, successfully stealing administrator API keys.
Why It Matters
This demonstration represents a watershed moment in cybersecurity research. For the first time, a credible expert has publicly acknowledged that AI has achieved superhuman performance in a critical security task.
The implications are far-reaching:
For Cybersecurity Professionals: Traditional manual code auditing may soon be insufficient. Security teams will need to integrate AI-powered vulnerability discovery tools or risk missing critical flaws that AI can easily identify.
For Software Developers: The discovery of a 21-year-old Linux vulnerability highlights how traditional review processes can miss subtle but serious flaws. AI-assisted code review may become essential for identifying complex vulnerabilities.
For Organizations: Companies relying on “security through obscurity” or assuming their code is secure because no vulnerabilities have been found may need to reassess their security posture.
Background
Nicolas Carlini brings exceptional credibility to this assessment. As a Research Scientist at Anthropic, he has:
- Published over 100 academic papers on AI safety and adversarial machine learning
- Accumulated more than 67,000 citations on Google Scholar
- Conducted foundational research on prompt injection attacks and AI alignment
- Previously worked at Google Brain on adversarial examples and model robustness
Carlini’s admission is particularly significant because he has spent years studying AI limitations and potential dangers. His acknowledgment that Claude has surpassed human expert capabilities in vulnerability discovery carries substantial weight in the security community.
The demonstration builds on Anthropic’s broader research into AI-powered security analysis. In related work, Anthropic’s AI agents successfully identified vulnerabilities in smart contracts worth approximately $4.6 million in simulated funds across 405 contracts deployed between 2020 and 2025.
What’s Next
Carlini predicts that AI vulnerability discovery capabilities will continue improving rapidly. This trend could fundamentally reshape cybersecurity:
Defensive Applications: Organizations may soon deploy AI systems to continuously audit their own codebases, potentially discovering vulnerabilities faster than human security teams.
Offensive Concerns: The same capabilities that help defenders could be misused by attackers to discover exploits at unprecedented scale and speed.
Regulatory Response: Governments may need to develop new frameworks for AI-discovered vulnerabilities, including mandatory disclosure requirements and defensive deployment guidelines.
Industry Standards: Software development practices may evolve to incorporate AI-powered security analysis as a standard part of the development lifecycle.
The cybersecurity community is now grappling with a fundamental shift: AI systems that can outperform human experts at finding the very vulnerabilities that threaten digital infrastructure.