Pith. sign in

REVIEW 19 cited by

AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2407.12784 v1 pith:GKZNGXLX submitted 2024-07-17 cs.LG cs.CRcs.IR

classification cs.LGcs.CRcs.IR
keywords knowledgeagentpoisonagentsbackdoormemorytriggeragentbases
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

LLM agents have demonstrated remarkable performance across various applications, primarily due to their advanced capabilities in reasoning, utilizing external knowledge and tools, calling APIs, and executing actions to interact with environments. Current agents typically utilize a memory module or a retrieval-augmented generation (RAG) mechanism, retrieving past knowledge and instances with similar embeddings from knowledge bases to inform task planning and execution. However, the reliance on unverified knowledge bases raises significant concerns about their safety and trustworthiness. To uncover such vulnerabilities, we propose a novel red teaming approach AgentPoison, the first backdoor attack targeting generic and RAG-based LLM agents by poisoning their long-term memory or RAG knowledge base. In particular, we form the trigger generation process as a constrained optimization to optimize backdoor triggers by mapping the triggered instances to a unique embedding space, so as to ensure that whenever a user instruction contains the optimized backdoor trigger, the malicious demonstrations are retrieved from the poisoned memory or knowledge base with high probability. In the meantime, benign instructions without the trigger will still maintain normal performance. Unlike conventional backdoor attacks, AgentPoison requires no additional model training or fine-tuning, and the optimized backdoor trigger exhibits superior transferability, in-context coherence, and stealthiness. Extensive experiments demonstrate AgentPoison's effectiveness in attacking three types of real-world LLM agents: RAG-based autonomous driving agent, knowledge-intensive QA agent, and healthcare EHRAgent. On each agent, AgentPoison achieves an average attack success rate higher than 80% with minimal impact on benign performance (less than 1%) with a poison rate less than 0.1%.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 19 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. ChronoMem: Version Control and Semantic Rollback for Large Language Model Agent Memory

    cs.CL 2026-07 conditional novelty 7.0 of 10

    Memory versioning with semantic rollback lets LLM agents behave counterfactually after later interactions, improving rollback-consistent QA and summarization.

  2. Ground Truth First: A Longitudinal Evaluation Instrument for Agent Memory, and the Tenure Crossover in Memory-Architecture Rankings

    cs.CL 2026-07 conditional novelty 6.0 of 10

    A ground-truth-first synthetic memory benchmark shows that agent-memory architecture rankings invert with history length: short-horizon leaders lose at nine weeks.

  3. What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents

    cs.LG 2026-07 conditional novelty 6.0 of 10

    KV-cache eviction, prompt compression, recurrent state bounding, and agent memory consolidation are unified as one rate-distortion problem with a shared lower bound, shared failure mode, and transferable mechanisms.

  4. Epistemic Bias Injection: Manipulating LLM Opinion via Selective Context Retrieval

    cs.CR 2025-11 conditional novelty 6.0 of 10

    Factually correct but one-sided passages injected into a RAG database can shift LLM answers toward an attacker's stance, and a PCA/KL-based filter (BiasDef) reduces that shift.

  5. LLM in the Middle: A Systematic Review of Threats and Mitigations to Real-World LLM-based Systems

    cs.CR 2025-09 conditional novelty 6.0 of 10

    A systematic review that categorizes LLM threats, severity scores, and mitigations across development and operation life cycles and multiple deployment scenarios.

  6. Adaptive Content Restriction for Large Language Models via Suffix Optimization

    cs.CL 2025-08 conditional novelty 6.0 of 10

    SOP appends an optimized suffix to prompts, reducing generation of user-specified restricted terms across several LLMs while keeping output quality close to baseline prompting methods.

  7. From Seed to Harvest: Augmenting Human Creativity with AI for Red-teaming Text-to-Image Models

    cs.LG 2025-07 conditional novelty 6.0 of 10

    Seed2Harvest expands 1,000 human adversarial prompts into 27,650 LLM-generated variants that keep roughly comparable unsafe-image trigger rates and add hundreds of new geographic contexts.

  8. Defining and Detecting the Defects of the Large Language Model-based Autonomous Agents

    cs.SE 2024-12 conditional novelty 6.0 of 10

    This study defines eight defect types for LLM-based agents and presents Agentable, a CPG-plus-LLM static analysis tool that detects them with reported precision of 88.79% and recall of 91.03%.

  9. How Well Do AI Systems Solve AP Physics? A Comparative Evaluation of Large Language Models on Algebra-Based Free Response Questions

    physics.ed-ph 2026-03 unverdicted novelty 5.0 of 10

    ChatGPT 4.1 mini, Gemini 2.5 Flash, Claude 4.0 Sonnet, and DeepSeek R1 average 82–92% on AP Physics 1/2 free-response questions but systematically fail spatial, visual, and conceptual tasks.

  10. Layer-Wise Perturbations via Sparse Autoencoders for Adversarial Text Generation

    cs.CL 2025-08 reject novelty 5.0 of 10

    Sparse autoencoder activation perturbation (SFPF) applied on top of existing jailbreak prompts raises attack success rate on Qwen3-32B, but with no defense evaluation and weak reproducibility.

  11. Commercial LLM Agents Are Already Vulnerable to Simple Yet Dangerous Attacks

    cs.LG 2025-02 conditional novelty 5.0 of 10

    Simple prompt-injection attacks against real-world LLM agents like Anthropic Computer Use, MultiOn, and ChemCrow succeed at leaking data and enabling harmful actions.

  12. Beyond Reward Hacking: Causal Rewards for Large Language Model Alignment

    cs.LG 2025-01 conditional novelty 5.0 of 10

    Causal reward modeling applies MMD regularization to RLHF reward models to make reward scores statistically independent of spurious features, reducing measured length, sycophancy, concept, and demographic biases in ex...

  13. The Chronos Vulnerability: A Taxonomy of Temporal Persistence and Memory-Based Deception in Agentic AI

    cs.AI 2026-07 conditional novelty 4.0 of 10

    The paper is a taxonomy of persistent memory-based attacks on AI agents, reinterpreting known attacks (MINJA, EchoLeak, sleeper agents) under the umbrella 'Chronos Vulnerability' and proposing a defense-in-depth architecture.

  14. A Whole New World: Creating a Parallel-Poisoned Web Only AI-Agents Can See

    cs.CR 2025-08 conditional novelty 4.0 of 10

    A website can identify AI agents by their digital fingerprints and serve them a poisoned hidden version of the page, hijacking their actions via indirect prompt injection.

  15. Evaluation and Benchmarking of LLM Agents: A Survey

    cs.LG 2025-07 conditional novelty 4.0 of 10

    A review that proposes a two-dimensional taxonomy for evaluating LLM agents and highlights enterprise-specific evaluation gaps.

  16. Large Language Models in Cybersecurity: Applications, Vulnerabilities, and Defense Techniques

    cs.CR 2025-07 conditional novelty 4.0 of 10

    A survey that maps LLM applications, vulnerabilities, and defenses across eight cybersecurity domains, but with significant citation and rigor problems.

  17. Securing Agentic AI: From Per-Action Checks to Trajectory Assurance

    cs.AI 2026-08 conditional novelty 3.0 of 10

    A vision paper organizing agentic AI security into eleven research directions, with the thesis that safety requires verifiable trajectory-level behavioral guarantees, not per-action checks.

  18. Report on NSF Workshop on Science of Safe AI

    cs.CY 2025-06 unverdicted novelty 2.0 of 10

    An NSF workshop report articulating a cross-disciplinary research agenda for designing and verifying safe, trustworthy AI systems.

  19. A Survey on Backdoor Threats in Large Language Models (LLMs): Attacks, Defenses, and Evaluations

    cs.CR 2025-02 conditional novelty 2.0 of 10

    A literature review that taxonomizes LLM backdoor attacks and defenses by model construction phase, with no new experimental results.

Pith tools