Honggfuzz
Feedback-guided fuzzer used across Google. Handles multi-core workloads, hardware tracing, and corpus minimization to keep bug finds steady.
github.com/google/honggfuzzAgent Security Architecture · Google DeepMind
Robert Swiecki
I lead technical architecture for Agent Security at Google DeepMind. I have spent 25+ years in systems security, from vulnerability research to building the defenses that run in production.
Hardened sandboxing and execution-isolation infrastructure to safely run untrusted code and frontier AI agents at scale, paired with automated red-teaming and automated fix-verification loops.
Runtime guardrails, data-exfiltration controls, and prompt-injection defenses across Gemini's agentic products.
Leadership & Impact
-
Hardened Agent Sandboxing & Execution Isolation
Architecting multi-layered containment and syscall mediation to safely run autonomous frontier agents and untrusted code execution environments at scale, integrating automated red-teaming with closed-loop fix verification.
-
Agentic Runtime Guardrails & Exfiltration Controls
Designed and deployed runtime safety boundaries, indirect prompt-injection defenses, and data-exfiltration controls across Gemini's agentic ecosystem.
-
Foundational AI Systems Security (through early 2026)
Led Google's AI/ML Systems Security group, directing security architecture for core foundation models and training infrastructure via confidential compute, model supply chain integrity, and proactive red teaming.
-
Building High-Performing Security Teams
Grew engineering teams across Zurich, Mountain View, and Warsaw, pairing systems specialists with ML experts and holding them to clear threat reduction targets.
-
Shipping Production Isolation Frameworks
Launched sandboxing, syscall mediation, and hardening layers that protect major Google workloads, including Sandboxed API for running risky code with least privilege.
-
Critical Vulnerabilities Eliminated Pre-Launch
Pushed fuzzing and review programs that uncovered critical flaws in OpenSSL, the Linux kernel, browsers, and internal stacks before they shipped.
Flagship Projects
NsJail
Linux namespace-based jail for running untrusted code, CTF tasks, and production prototypes. Ships with flexible seccomp policies through the kafel DSL.
github.com/google/nsjailSandboxed API & Sandbox2
Composable runtime that lets teams execute high-risk code inside hardened compartments with low integration effort. Runs in latency-sensitive services and ML pipelines.
github.com/google/sandboxed-apiRecognition & Community
-
Pwnie Award Nominations
Recognized in 2016 and 2017 for AMD microcode privilege escalation work and for uncovering the first OpenSSL flaw rated 'critical'.
-
Dragon Sector CTF Founder
Co-founded and mentor Dragon Sector, a long-running competitive security team focused on new exploitation techniques and tooling.
-
Practical Reverse Engineering
Co-authored the Polish edition of 'Practical Reverse Engineering', turning advanced reversing techniques into practical guidance.