
co-weight-drift
PyTorch reference implementation of a fast adversarial training method that prevents catastrophic overfitting via coupled objectives and an EMA-based…

PyTorch reference implementation of a fast adversarial training method that prevents catastrophic overfitting via coupled objectives and an EMA-based…

Research code for measuring jailbreak attack success rates (ASR and stepwise ASR) against LLMs, with implementations of attacks such as GCG and…

Code, derived data and results for: Evaluating Baseline Detectors of Hijacked LLM Agents Under Rare Attacks and Distribution Shift

Research code for Rubric-Induced Preference Drift (RIPD): evolutionary rubric search, benchmark-preserving selection, and DPO policy misalignment…

Research codebase for privacy evaluation of quantized neural networks: trains target/shadow models under PTQ/QAT and runs membership-inference…

CyberRAG — RAG 100% local (Ollama) sobre apuntes y libros de ciberseguridad: recuperación híbrida, respuestas con citas, asistente de pentest por…

Research pipeline for detecting latent indirect prompt-injection exposure signals in agentic LLMs via hidden-state probing, including trace…

Research code reproducing multi-turn LLM jailbreak experiments (FITD, MRCJ, ActorAttack, X-Teaming) from the SoK intent-oriented systematization…

Detects LLM context-leakage attacks by training lightweight behavior probes on log-probabilities, with vLLM offline/server detection pipelines.

Collects multi-source security logs from reproducible Docker attack scenarios, builds attack chains, and provides the SecLog-Bench-216 benchmark for…

Pre-execution action-auditing defense that detects and masks indirect prompt injection in tool-using LLM agents using embedding retrieval and…

Jailbreak vulnerability across early-exit depths LLMs. Published at NeurIPS 2026 in the FLMSec workshop

Deterministic benchmark and evaluation harness for persistent-memory poisoning in LLM agents, with paired clean/poisoned cases, LangChain/LlamaIndex…

LLM-driven agentic group shilling attack framework that manipulates black-box collaborative-filtering recommender rankings using adaptive multi-role…

ML-based detection of Zombie ZIP archive header evasion attacks (CVE-2026-0866)

Simulation code and results for a private, verifiable, Byzantine-robust federated intrusion detection protocol using secret sharing, Merkle audits…

Research framework for detecting and removing backdoors in language models, combining feature-based anomaly detection with LoRA-based unlearning…

Adversary-resilient deep learning architecture for secure 5G indoor localization, combining CNN and multi-head attention to defend against signal…