Valentina Palmiotti aka Chompie y Sam Lovejoy I Lado B Ekoparty 2025

### Overview In this interview from the 'Lado B' track at Ekoparty 2025 in Buenos Aires, security researchers Valentina Palmiotti (Chompie) and Sam Lovejoy discuss their research into attacking Nvidia drivers. Motivated by the growing industry focus on AI and new targets at Pwn2Own, the researchers explored binary exploitation in kernel drivers that support container-accelerated AI workloads.

### Key Topics - **Target Selection and Motivation**: The researchers chose Nvidia due to company-wide pressure to pursue AI-related topics, the lack of existing public binary exploitation research on Nvidia products, and new targets introduced at Pwn2Own. - **Container Isolation and Cloud Infrastructure**: They examined how kernel driver flaws impact cloud architectures. Although the container runtime itself appeared relatively hardened, major cloud platforms rely heavily on Nvidia Container Tools and container boundaries for customer isolation. - **Role of LLMs in Security**: The discussion touched on whether AI will displace security professionals, emphasizing that models act more as copilots for pattern recognition, vulnerability discovery, and targeted tasks rather than fully autonomous exploit developers. - **Advice for Beginners**: The speakers offered practical tips for entering kernel security research and binary exploitation.

### Technical Details - **Container Escape via Drivers**: The research focused on escaping containers running kernel-accelerated AI workloads by targeting low-level Nvidia kernel drivers rather than the higher-level container runtime. - **Cloud-Level Vulnerability Scope**: The researchers noted that finding bugs took only three weeks once testing began. Because major cloud providers (such as Google, DigitalOcean, and Replicate) rely on Nvidia Container Tools and container isolation between distinct clients, a vulnerability inside the kernel driver escalates to an isolation failure across multi-tenant cloud environments. - **Ecosystem Developments**: The researchers stated their exploit was among the first public driver-level exploits in this stack. Shortly after their work, Quarkslab also published research detailing driver-level exploitation. - **AI Specialization**: Rather than expecting an LLM to manage an end-to-end exploitation chain, Chompie highlighted the effectiveness of fine-tuning models for narrowly scoped tasks, such as automatically recalculating and updating exploit offsets when new software versions are released.

### Takeaways - **Kernel Exploitation Practice**: Beginners in vulnerability research are encouraged to replicate existing public exploits, rewrite them line-by-line to understand internal mechanics, and use learning platforms like Google's Kernel CTF. - **Human-in-the-Loop AI**: Job displacement is unlikely in the next five or more years. AI tools will primarily assist human practitioners by clearing low-hanging fruit and speeding up pattern identification and debugging. - **Personal Milestones**: Valentina highlighted her return to Ekoparty, fulfilling a long-time goal of delivering a talk in Spanish after speaking there initially in 2021.