MemoryDojo benchmarks poisoning attacks on LLM agent memory
Benchmark with 155 trajectories demonstrates all tested memory systems are susceptible to poisoning.
Stories per day · 30 days
4 RSS
Relatedbackdoor4ai-security2rag2retrieval2agent1ai-misuse1asv1backdoors1benchmarks1chatgpt1clean-label1clip1data-sanitization1defense1defenses1distributed-ml1ember1face-recognition1
13 stories, newest first · one card per story
Stories tagged #poisoning
Lawsuit: ChatGPT helped plan husband's murder via poisoning
ChatGPT generated personalized poisoning instructions using oleander and other drugs; OpenAI banned the user and shared logs with law enforcement.
Three defenses harden face recognition
Three tricks detect poisoned identities at training and filter inputs at inference.
Pop Goes the Verifier poisons voice liveness detection
Phoneme-targeted poisoning using FastSpeech 2 degrades ASVSpoof 2019 voice liveness detection accuracy with stealth against waveform inspection.
Survey maps security risks in distributed and federated learning
Federated learning is treated as a high-risk case, with defenses compared from aggregation to trusted execution.
Paper says RAG poison filters rely on construction shortcuts, not content
Stripping PoisonedRAG's question prefix drops reimplemented RAGuard from AUROC 0.953 to 0.758 and raises its miss rate to 0.763.
Poisoning the Neighborhood: A Topology-Aware Topic-Level Attack on Large-Scale HNSW-based RAG Systems
Researchers demonstrate a Hub-Satellite poisoning attack on HNSW RAG with 31.29% success on two million Wikipedia documents.
Trigger Value Selection in Clean-Label Malware Backdoors: An Effectiveness-Detectability Analysis
Researchers show trigger-value choice changes success and detectability of clean-label backdoors in malware classifiers.
Clean Data Can Still Carry Backdoors: Support-Persistent Backdoors for Model Reuse
DAR plants support-persistent backdoors that stay above 96 percent success after clean-data distillation.
Analyzing and Mitigating Poisoning Attacks on Flow-Based Intrusion Detection
Researchers backdoor flow-based IDS models with 0.001% poisoned data and propose mitigations.
Bounded Provisional Visibility: Controlling Poisoning Exposure in Continuously Ingested RAG Vector Stores
A fail-closed protocol limits how long unvetted RAG documents remain retrievable during continuous ingestion.
Online AutoML: Evaluating Poisoning Attacks on Adversarial Training Defense Strategy in IoT Networks
Study tests adversarial training against label-flip and noise poisoning of streaming AutoML models for IoT.
Poster: FedWM-Guard: Thwarting Imagination Poisoning in Federated World Model-based Autonomous Driving
A poster proposes FedWM-Guard to stop imagination poisoning of federated autonomous-driving world models.