Agent Security Research Papers

Contributions Welcome License GitHub Pages

We aim to systematically track agent-security research across security and software engineering conferences and journals. Because AI venues publish a much larger volume of papers, we collect relevant AI work as comprehensively as possible and continuously expand the bibliography. The collection covers prompt injection, tool and action security, memory poisoning, privacy leakage, agent misuse, benchmarks, red teaming, and system-level defenses. Contributions and corrections are welcome.

Contents

2026 2025 2024
Benchmarks and Evaluation Prompt Injection Attacks Defenses and Secure Design Tool and Action Security Memory, RAG, and Privacy Agent Misuse and Red Teaming Preprints

All Papers

2026

AI conference entries retain their official presentation labels.

ICSE

FSE

ASE

ISSTA

IEEE TSE

USENIX Security

NDSS

IEEE S&P

CCS

NeurIPS

ICLR

ICML

ACL

EMNLP

2025

ASE

ICSE

NDSS

USENIX Security

NeurIPS

ICLR

ICML

KDD

ACL

NAACL

EMNLP

CCS

2024

ASE

IEEE TDSC

NeurIPS

ICML

ICLR

ACL

USENIX Security


Papers by Topic

Benchmarks and Evaluation

Benchmarks, test environments, and empirical evaluations that measure agent security, harmful task completion, or deployment risk.

General Agent Security

Prompt Injection Benchmarks

Prompt Injection Attacks

Attacks that place malicious instructions in external content, tool streams, retrieval corpora, or multi-step agent context.

Indirect and Context-Aware Attacks

Retrieval and Tool-Stream Attacks

Defenses and Secure Design

Model-level, runtime, and architectural defenses that preserve user intent and constrain agent actions.

Instruction and Intent Defenses

Runtime and Graph-Based Defenses

Tool and Action Security

Unsafe tool selection, taint propagation, permission control, and action validation in tool-using agents.

Plugins, APIs, and Agent Frameworks

Tool Invocation and Taint

Web and Environment Interaction

Memory, RAG, and Privacy

Attacks and defenses involving persistent memory, knowledge bases, private data, or cross-tool information flows.

Memory and Knowledge-Base Poisoning

Privacy Leakage

Agent Misuse and Red Teaming

Research on harmful task execution, offensive agents, automated red teaming, and defenses against malicious agents.

Harmful and Offensive Agents

Jailbreaks, Backdoors, Collusion, and Reward Hacking


Preprints


Contributing

Contributions are welcome. Please open an issue or pull request to:

Use the following entry format and prefer an official proceedings, publisher, ACL Anthology, OpenReview, or arXiv link:

- [Paper Title](official-paper-url)

License

The organization and documentation in this repository are licensed under CC BY 4.0. Individual papers remain copyrighted by their authors and publishers.