3 rd ACM Workshop on
Large AI Systems and Models with Privacy and Security Analysis
November 15, 2026 — The Hague, The Netherlands
co-located with the 33rd ACM Conference on Computer and Communications Security

Keynotes

Rethinking Cybersecurity for AI Agents That Do What You Say, Not What You Mean

Associate Professor Fabio Pierazzi, University College London

Fabio Pierazzi is an Associate Professor in Computer Science at University College London (UCL) and a member of the UCL Information Security Group.

His research focuses on leveraging AI to enhance systems security, particularly in environments where attackers rapidly adapt to new defenses. His key research interests include measuring, mitigating, and preventing concept drift in ML-based detection systems; improving adversarial robustness while accounting for problem-space constraints, such as preserving code semantics; and explaining the behavior of security systems from both attacker and defender perspectives.

He places strong emphasis on the practicality of AI-driven security solutions and understanding their real-world implications and limitations, actively collaborating with industry partners. His research primarily covers malware analysis and network traffic analysis, with growing interests in understanding the inner workings of AI and ML models to improve their trustworthiness across broader security applications.

We now live in a world of deployed agents that have moved from answering questions to executing complex tasks with real credentials on real systems. We have already seen this go wrong: agents cheat to finish a task, escape their sandboxes, and compromise external services they were never asked to touch. Ask a person for a coffee and they buy one or make one; ask an AI agent, and it might steal one, because nothing in the request ruled it out.

This talk explores what specifying agent behaviour would actually take. Towards an answer, the talk will delve deeper into a far more constrained scenario: Deep Reinforcement Learning (DRL) agents, where state, actions, transitions, and rewards are all written down explicitly. Even here, proper specification, modelling and evaluation are hard. Drawing on our USENIX Security 2026 paper, this talk will present eleven pitfalls in DRL for cybersecurity tasks, spanning training, evaluation, and deployment, with recommendations for each. The talk will close by looking ahead at the open challenges this raises for specifying agent behaviour more broadly.

A Neuron a Day Keeps Rogue Agents Away - From Model Editing to Secure Autonomous AI

Professor Ahmad-Reza Sadeghi, Technical University Darmstadt, Germany

Ahmad-Reza Sadeghi is a Full Professor of Computer Science at the Technical University of Darmstadt, Germany, where he heads the System Security Lab. Since 2012 Prof. Sadeghi has established a long-term cooperation with Intel. It has already emerged in several Collaborative Research Centers on various topics, such as Secure Computing in Mobile and Embedded Systems, Autonomous and Resilient Systems, and Private AI. Moreover, he also established the Open Lab for Sustainable Security and Safety (OpenS3 Lab) with Huawei in 2019.

received his Ph.D. in Computer Science with a focus on Cryptography from the University of Saarland, Germany. Before academia, he worked for several years in the Research and Development of the Telecommunications industry, amongst others, Ericsson. He has been leading and involved in many national and international research and development projects in the design and implementation of Trustworthy Computing Platforms, Hardware-assisted Security, IoT Security and Privacy, Applied Cryptography, and Trustworthy AI.

Sadeghi has been serving as General or Program Chair and Program Committee member of major Information Security and Privacy and Design and Automation venues, such as ACM CCS, IEEE Security & Privacy, NDSS, USENIX Security, DAC, DATE, and ICCAD. He was Editor-In-Chief of IEEE Security and Privacy Magazine. Ahmad served on several editorial boards, such as ACM Transactions on Information & System Security (TISSEC), as Guest editor of the IEEE TCAD, ACM Books, and ACM DIOT. He is on the editorial board of ACM TODAES and ACM DTRAP.

As large language models evolve from answering questions to powering autonomous agents, unsafe model behavior can propagate through planning, memory, and tool use, turning a single flawed decision into potentially harmful real-world actions. Securing the underlying model is therefore essential for trustworthy AI agents.

In this talk, we explore targeted model editing and neuron-level interventions as a promising approach to building lightweight, efficient, and inherently safer AI systems. After briefly reviewing the limitations of existing safety approaches, we show how modifying a small subset of neurons can improve model safety without costly retraining. We discuss applications of this approach in post-hoc safety hardening, particularly after domain-specific fine-tuning, e.g., in medical consultation and financial reasoning.

Finally, we discuss emerging research directions that build on these internal interventions to extend model-level security toward securing the decisions and actions of autonomous AI coding agents.

Programme

08:30 Registration and welcome coffee
09:00-09:10 Opening and Best Paper Award
Keynote 1
09:10-10:00 Rethinking Cybersecurity for AI Agents That Do What You Say, Not What You Mean
Fabio Pierazzi, Associate Professor, University College London
10:00-10:30 Coffee break and posters
Session 1
10:30-10:45 Upstream Supply Chain Security in MCP: Structural Limits of SAST and Deployment-Side Defense
Youna Jung, Uryaswi Bhowmick (Northeastern University)
10:45-11:00 From Guardrail Leak to Agentic Infrastructure Compromise: Lateral Movement Across an LLM-Orchestrated Kubernetes Platform
Emre Süren, Rebecca Ehrnrooth From, Alexander Timsäter (KTH Royal Institute of Technology)
11:00-11:15 Beyond Attack Success Rate: Examining Trigger Leakage in Vision-Language Agentic Systems
Jiamin Chang, Salil Kanhere, Piotr Koniusz (University of New South Wales), Minhui Xue (CSIRO and Adelaide University), Hammond Pearce (University of New South Wales)
Session 2
11:15-11:30 SoK: The Security-Safety Continuum of Multimodal Foundation Models (🏆Best Paper Award)
Ruoxi Sun (CSIRO), Jiamin Chang (University of New South Wales and CSIRO), Hammond Pearce (University of New South Wales), Chaowei Xiao (Johns Hopkins University), Bo Li (University of Illinois Urbana-Champaign), Qi Wu (Adelaide University), Surya Nepal (CSIRO), Jason Xue (CSIRO and Adelaide University)
11:30-11:45 Design Exploration for Context-Based Access Control of Vision-Language-Action Model-Based Embodied Agents
Krutyanjay Shinde, Dongha Kim, Hokeun Kim (Arizona State University)
11:45-12:00 Position: Agents Orchestrate, Policy Enforces - A Deterministic Tool-Use Control Pattern for Agentic Assistants over Sensitive EEG Data
Sofio Katamadze (The University of Georgia), Ankur Chattopadhyay (Northern Kentucky University)
12:00-12:15 Private LLM Inference: Optimizing Tricycle
Zaira Pindado, Hugo Sanz-González, Thomas Spendlhofer, Antonio J. Peña (Barcelona Supercomputing Center)
12:15-13:30 Lunch
13:30-14:00 Poster session
Keynote 2
14:00-15:00 A Neuron a Day Keeps Rogue Agents Away -- From Model Editing to Secure Autonomous AI
Ahmad-Reza Sadeghi, Professor, Technical University Darmstadt, Germany
15:00-15:30 Coffee break and posters
Session 3
15:30-15:45 Embodied AI Must Treat Visual Bias as a Safety Primitive, Not a Robustness Penalty
Enguang Liu (Nanjing University of Science and Technology), Zhigang Lu (Western Sydney University), Si-Yuan Liang (National University of Singapore), Liming Lu, Xiyu Zeng, Jiasheng Cai (Nanjing University of Science and Technology), Xiaochun Cao (Sun Yat-Sen University), Aishan Liu (Beihang University), Shuchao Pang (Nanjing University of Science and Technology)
15:45-16:00 Delta-Weight Quantization Preserves Alignment in Large Language Models
Yule Liu, Zhen Sun (The Hong Kong University of Science and Technology Guangzhou), Xinlei He (Wuhan University)
16:00-16:15 Safer Content or Firmer Refusals? A Hybrid Perturbation Defense for Alignment under Harmful Fine-tuning
Muhammad Zeeshan Akram, Mufid Kamel Marican, Anvesh Reddy Yenugu, Ali Zain Kaimkhani, Minghong Fang (University of Louisville)
16:15-16:30 Does the Unsafe Gradient Survive a Conversation? On the Fragility of Gradient-Based Jailbreak Detection in Multi-Turn Dialogue
Omar Sheta, Rinku Deuja, Hadi Masoudi, Minghong Fang (University of Louisville)
16:30-16:40 Closing

Accepted Papers

Agentic System Security: Infrastructure, Supply Chain & Attack Surfaces
Upstream Supply Chain Security in MCP: Structural Limits of SAST and Deployment-Side Defense
Authors : Youna Jung, Uryaswi Bhowmick (Northeastern University)
From Guardrail Leak to Agentic Infrastructure Compromise: Lateral Movement Across an LLM-Orchestrated Kubernetes Platform
Authors : Emre Süren, Rebecca Ehrnrooth From, Alexander Timsäter (KTH Royal Institute of Technology)
Beyond Attack Success Rate: Examining Trigger Leakage in Vision-Language Agentic Systems
Authors : Jiamin Chang, Salil Kanhere, Piotr Koniusz (University of New South Wales), Minhui Xue (CSIRO and Adelaide University), Hammond Pearce (University of New South Wales)
Alignment Robustness under Fine-Tuning, Quantization & Multi-Turn Attacks
Safer Content or Firmer Refusals? A Hybrid Perturbation Defense for Alignment under Harmful Fine-tuning
Authors : Muhammad Zeeshan Akram, Mufid Kamel Marican, Anvesh Reddy Yenugu, Ali Zain Kaimkhani, Minghong Fang (University of Louisville)
Delta-Weight Quantization Preserves Alignment in Large Language Models
Authors : Yule Liu, Zhen Sun (The Hong Kong University of Science and Technology Guangzhou), Xinlei He (Wuhan University)
Does the Unsafe Gradient Survive a Conversation? On the Fragility of Gradient-Based Jailbreak Detection in Multi-Turn Dialogue
Authors : Omar Sheta, Rinku Deuja, Hadi Masoudi, Minghong Fang (University of Louisville)
Multimodal & Embodied AI: Safety Foundations
SoK: The Security-Safety Continuum of Multimodal Foundation Models
Authors : Ruoxi Sun (CSIRO), Jiamin Chang (University of New South Wales and CSIRO), Hammond Pearce (University of New South Wales), Chaowei Xiao (Johns Hopkins University), Bo Li (University of Illinois Urbana-Champaign), Qi Wu (Adelaide University), Surya Nepal (CSIRO), Jason Xue (CSIRO and Adelaide University)
Embodied AI Must Treat Visual Bias as a Safety Primitive, Not a Robustness Penalty
Authors : Enguang Liu (Nanjing University of Science and Technology), Zhigang Lu (Western Sydney University), Si-Yuan Liang (National University of Singapore), Liming Lu, Xiyu Zeng, Jiasheng Cai (Nanjing University of Science and Technology), Xiaochun Cao (Sun Yat-Sen University), Aishan Liu (Beihang University), Shuchao Pang (Nanjing University of Science and Technology)
Policy Enforcement, Access Control & Privacy-Preserving Deployment
Position: Agents Orchestrate, Policy Enforces — A Deterministic Tool-Use Control Pattern for Agentic Assistants over Sensitive EEG Data
Authors : Sofio Katamadze (The University of Georgia), Ankur Chattopadhyay (Northern Kentucky University)
Design Exploration for Context-Based Access Control of Vision-Language-Action Model-Based Embodied Agents
Authors : Krutyanjay Shinde, Dongha Kim, Hokeun Kim (Arizona State University)
Private LLM Inference: Optimizing Tricycle
Authors : Zaira Pindado, Hugo Sanz-González, Thomas Spendlhofer, Antonio J. Peña (Barcelona Supercomputing Center)

Call for Papers

Important Dates

  • Paper submission deadline: July 19, 2026, 11:59 PM (all deadlines are AoE, UTC-12)
  • Acceptance notification: August 24, 2026
  • Camera ready due: September 7, 2026
  • Workshop day: November 15, 2026

Overview

The trajectory of AI has fundamentally shifted in 2026. We are moving beyond standalone Large AI Models (LAMs) to deeply integrated, autonomous Agentic AI systems capable of long-term planning, external tool execution, and local shell access. While these agentic frameworks promise to revolutionize critical infrastructure management, they simultaneously introduce a massive system-level attack surface. In this new paradigm, traditional empirical defenses are no longer sufficient.

Building upon the resounding success and immense community impact of LAMPS'24 and LAMPS'25, which consistently drew over 80 attendees and highly competitive submissions, LAMPS'26 is dedicated to tackling this exact frontier. As a recognized catalyst for cutting-edge AI security research, we place a spotlight on mathematically grounded defenses this year. We explicitly bridge Agentic Security with rigorous Formal Verification and Advanced Cryptography to secure high-stakes autonomous systems, effectively setting the research agenda for the future of secure AI deployment.

Topics of Interest

Topics of interest include (but are not limited to):

Track 1: Agentic System Security & Formal Verification (Featured Theme)

  • Formal Guarantees for Tool-Use: Applying formal verification to bound agent actions, mathematically proving safe API execution and preventing sandbox escapes.
  • Skill Exploitation & IPI: Defending against the malicious manipulation of agent "skills" (e.g., shell execution, web browsing) via indirect prompt injection.
  • The "Double Agent" Risk: Mitigating threats where personalized agents autonomously execute unauthorized actions without human-in-the-loop oversight.
  • Agent-to-Agent (Multi-Agent) Vulnerabilities: Investigating emergent interaction risks such as agent collusion, adversarial manipulation of peer agents (e.g., cross-agent prompt injection), and systemic cascading failures.
  • Agent-Environment and Tool Interaction Risks: Examining compounding vulnerabilities when agents autonomously interact with external environments, APIs, and untested data sources, leading to unauthorized traversal or payload execution.

Track 2: Cryptographic Defenses & Privacy-Preserving AI

  • Cryptographic Agent Memory: Leveraging Fully Homomorphic Encryption (FHE) and Secure Multi-Party Computation (SMPC) for encrypted agent states and secure collaborative reasoning.
  • Verifiable Agent Actions: Utilizing Zero-Knowledge Proofs (ZKPs) to audit autonomous agent workflows without revealing underlying prompts or sensitive user data.
  • Advanced Data Protection: Data anonymization, synthetic data generation, and IP protection in model weights via cryptographic watermarking.
  • Countermeasures against advanced adversarial attacks, data poisoning, and model extraction in continuous-learning systems.

Track 3: Secure LAMs for Critical Infrastructure

  • AI-driven automated security orchestration, penetration testing, and incident response (SOAR).
  • Vulnerability assessments of LAMs deployed in industrial control systems and cyber-physical environments.
  • Collaborative AI for secure, cross-border global cyber threat intelligence sharing.

Track 4: Human-Centric AI Security & Governance

  • Human-Agent Interaction Exploitation: Addressing the risks of deceptive alignment, advanced social engineering by autonomous agents, and vulnerabilities stemming from human over-reliance on agentic workflows.
  • Human factors in adversarial vulnerabilities and social engineering via autonomous AI assistants.
  • Combatting deepfakes to preserve trust in digital communications.
  • Explainability, fairness, and transparent decision-making in high-stakes autonomous systems.
  • Frameworks for responsible AI governance and participative stakeholder engagement.

Submission Guidelines

Submitted papers must not substantially overlap with papers that have been published or simultaneously submitted to a journal or a conference with proceedings.

  • Short Papers: These papers should present concise and focused contributions, such as preliminary research findings, novel ideas with early evidence, or case studies relevant to the aforementioned topics of interest. Submissions must be up to 4 pages of body text in the ACM double-column format. Short papers must offer a clear and well-motivated contribution, even if the work is at an early stage, and should be of interest to the research community.
  • Research Papers: These papers should present new work, evidence, or ideas related to aforementioned topics of interest. Submission must be up to 8 pages of body text in the ACM double-column format, excluding well-marked references and appendices, and at most 10 pages. Research papers must be well-argued and worthy of publication and citation, on one of the topics listed above.
  • Systematization of knowledge (SoK) Papers: These papers should either consolidate and clarify ideas in a major research area within secure and trustworthy machine learning or provided compelling evidence to support or challenge long-held beliefs in such areas. Submission must be up to 8 pages of body text in the ACM double-column format, excluding well-marked references and appendices, and at most 10 pages. SoK papers must include "SoK:" at the beginning of their title.
  • Position Papers: These papers should cover broader issues and visions related to aforementioned topics of interest, including open challenges, technical perspectives, educational aspects, societal impact, or notable research results. Submissions must be very well-argued and consist of at most 4 pages of body text in the ACM double-column format, excluding well-marked references and appendices, and at most 5 pages in total. Position papers must include "Position:" at the beginning of their title.

Submission Site

Submission link: https://ccs26-lamps.hotcrp.com

Only PDF files will be accepted. Submissions not meeting these guidelines risk rejection without consideration of their merits. Authors of accepted papers must guarantee that one of the authors will register and present the paper at the workshop. Proceedings of the workshop will be available on a CD to the workshop attendees and will become part of the ACM Digital Library.

The archival papers will be included in the workshop proceedings. Due to time constraints, accepted papers will be selected for presentation as either talk or poster based on their review score and novelty. Nonetheless, all accepted papers should be considered as having equal importance.

Authors are responsible for obtaining appropriate publication clearances. Attendance and presentation by at least one author of each accepted paper at the workshop are mandatory for the paper to be included in the proceedings.

For any questions, please contact one of the PC co-chairs Xiaoyu Xia: xiaoyu.xia@rmit.edu.au.

Committee

PC Chairs

Web/Publication Chair

Organizing Committee