Paul Kassianik
Principal AI Security Researcher / AI Security Lead, Foundation AI (Cisco). Based in San Francisco. Building AI systems for cybersecurity.
About
Now: leading research and pretraining for the Foundation-Sec model family at Cisco — open-weight LLMs purpose-built for security operations (base, instruct, reasoning).
Recent: automated red-teaming (Tree of Attacks, NeurIPS ‘24), prompt-injection defenses, and risk evaluation of frontier reasoning models.
Earlier: code-generation LLMs and industrial time-series ML (Merlion) at Salesforce Research.
Experience
Cisco — Foundation AI (2024 – Present) Principal Researcher. Joined via Cisco’s acquisition of Robust Intelligence. Leads pretraining, data, and evaluation for the Foundation-Sec model family (base, instruct, reasoning). Architects AI-Firewall research and production guardrails.
Robust Intelligence (2023 – 2024, acquired by Cisco) Principal Researcher. Built synthetic jailbreak frameworks that integrated new attack vectors into products within 72 hours; improved prompt-injection detection rates by ~20%.
Salesforce Research (2021 – 2023) Researcher. Code-generation large language models and industrial time-series analysis. Core contributor to Merlion, Salesforce’s open-source time-series ML library.
Selected publications
- [FEATURED] Tree of Attacks: Jailbreaking Black-Box LLMs Automatically — Mehrotra, Zampetakis, Kassianik, Nelson, Anderson, Singer, Karbasi. NeurIPS 2024 / arXiv:2312.02119.
- [FEATURED] Llama-3.1-FoundationAI-SecurityLLM-Base-8B Technical Report (lead author) — Kassianik, Saglam, Chen, Nelson, Vellore, et al. arXiv preprint arXiv:2504.21039, 2025.
- Merlion: End-to-End Machine Learning for Time Series — Bhatnagar, Kassianik, Liu, Lan, Yang, Cassius, Sahoo, et al. Journal of Machine Learning Research, 24(226), 2023 / arXiv:2109.09265.
- Adversarial Reasoning at Jailbreaking Time — Sabbaghi, Kassianik, Pappas, Singer, Karbasi, Hassani. ICML 2025 / arXiv:2502.01633.
- Llama-3.1-FoundationAI-SecurityLLM-8B-Instruct Technical Report — Weerawardhena, Kassianik, Nelson, Saglam, Vellore, et al. arXiv preprint arXiv:2508.01059, 2025.
- Large Language Models Encode Semantics in Low-Dimensional Linear Subspaces — Saglam, Kassianik, Nelson, Weerawardhena, Singer, Karbasi. IJCNLP-AACL 2025 / arXiv:2507.09709.
- Extracting Memorized Training Data via Decomposition — Su, Vellore, Chang, Mura, Nelson, Kassianik, Karbasi. arXiv preprint arXiv:2409.12367, 2024.
- Capability-Based Scaling Laws for LLM Red-Teaming — Panfilov, Kassianik, Andriushchenko, Geiping. ICML 2025 Workshop on Reliable and Responsible Foundation Models / arXiv:2505.20162.
- Llama-3.1-FoundationAI-SecurityLLM-Reasoning-8B Technical Report — Yang, Li, He, Priyanshu, Saglam, Kassianik, et al. arXiv preprint arXiv:2601.21051, 2026.
- A Framework for Rapidly Developing and Deploying Protection Against Large Language Model Attacks — Swanda, Chang, Chen, Burch, Kassianik, Berlin. CAMLIS 2025 / arXiv:2509.20639.
See Google Scholar for the full, up-to-date list and current citation counts.
Patents
- Systems and methods for jailbreaking black-box large language models — Mehrotra, Kassianik. U.S. Patent App. 18/957,525, 2025.
Open source
- Foundation-Sec-8B — 8-billion-parameter open-weight base model specialized for cybersecurity. Continued pretraining of Llama-3.1-8B on ~5.1B tokens of in-house-curated security data. Lead author / research lead.
- Merlion — Salesforce’s open-source end-to-end machine-learning framework for time series (~4.5k stars on GitHub). Core contributor.
Writing
Author at the Cisco Security Blog:
- [FEATURED] Evaluating Security Risk in DeepSeek and Other Frontier Reasoning Models — January 2025
- Accelerate Security Operations with Cisco’s New Security-Tuned Model — February 2026
- Using AI to Automatically Jailbreak GPT-4 and Other LLMs in Under a Minute — December 2023
Education
B.A. Applied Mathematics, University of California, Berkeley — Computer Science focus.