Fraud control gets a blockchain audit trail here, with synthetic telecom/IoT requests, federated meta-learning, and QLoRA LLMs.
trending30
01 02 PromptGraph: Graph-Guided Prompt Sanitization for Balancing Privacy and Utility in LLM Inference arxiv.orgCan a graph over prompt spans scrub PII without torching LLM utility?03 Can an LLM write EasyCrypt proofs, and this repo ships the MCP workflow, benchmarks, and replay logs to find out.04 TRACE: A Two-Channel Robust Attribution Watermark via Complementary Embeddings for LLM-Agent Trajectories arxiv.orgTRACE watermarks LLM-agent trajectories with two keyed channels, trying to survive log deletion and rewriting, which is the annoying part.05 Can you trace a poisoned code completion back to the fine-tuning sample that taught it? CodeTracer says yes, mostly.06 Mechanistic jailbreak analysis with paired clean/attacked attribution graphs, finally doing more than waving at prompt gradients.07 Large Language Models (LLMs) and Generative AI in Cybersecurity and Privacy: A Survey of Dual-Use Risks, AI-Generated Malware, Explainability, and Defensive Strategies arxiv.orgHow bad is the dual-use mess? This survey maps LLM cyber use from phishing and malware to explainability and defenses.08 24 bits of watermark in generated code, spread across names and semantics, with proofs and Reed-Solomon keeping score.09 Code-Level Cost Function Generation for Spatial Image Steganography Using RAG-Enhanced Large Language Models arxiv.orgLLMs are now writing steganography cost functions, with RAG and a self-evolving knowledge base in the loop.10 Arxiv benchmark for polymorphic Sybil poisoning in RAG, with forced exposure and a 4-way failure taxonomy instead of yet another toy attack.11 HeRo watermarks LLM output with hierarchical vocab routing, so verifiers only see the metadata they're cleared for.12 RustMizan: A Compilable, Contamination-Aware Benchmarking Framework for Rust Vulnerabilities arxiv.orgRustMizan is a Rust vuln benchmark with compilable variants, CWE labels, and contamination tests, since folklore needed CI.13 Can you fingerprint the embedding model from unordered retrieval results? Apparently yes, and the paper tests rerankers and RAG too.14 Paper proposes EvoVuln, which evolves smart-contract vuln rules from a few labels instead of touching model weights.15 End-to-end confidential AI for CPU+GPU TEEs, with attestation and benchmarks, not the usual single-box secure inference demo.16 Words Speak Louder Than Code: Investigating Cognitive Heuristics in LLM-Based Code Vulnerability Detection arxiv.orgPaper on heuristic bias in LLM vuln detectors, plus a black-box attack that flips verdicts without touching code. Charming.17 Client-side proxy swaps PII for type-matched surrogates, then stitches the originals back, since plain redaction wrecks utility.18 Fine-tuning fewer steps, not DP-SGD, does most of the memorization reduction; HMAC pseudonymization helps on identifiers.19 On the Internet, Nobody Knows You're an LLM Bot: Unmasking Web Agents with Multi-Layer Fingerprinting arxiv.orgThey beat CAPTCHAs, then get fingerprinted anyway, which is the paper's point about LLM web agents.20 FlipGuard: Defending Large Language Models Against Quantization-Conditioned Backdoor Attacks arxiv.orgPaper on defending LLMs from quantization-triggered backdoors with FlipGuard, because even weights need a seatbelt.21 ToolPrivacyBench, a benchmark for purpose-bound privacy in tool-using LLM agents, because apparently prompts needed audits.22 Paper on a malicious federated QA aggregator planting data-free backdoors from gradients, because trust is apparently optional.23 Credentials leak from third-party agent skills more often than you'd hope, per a large-scale arXiv study.24 How does Claude's sandboxed memory actually work, and where do the creds and plaintext leak out?25 TRACE flags poisoned RAG corpora by finding repeated high-influence tokens and checking if they steer answers, sadly enough.26 Survey of privacy risks in LLM agents, from retrieval to memory to prompt injection, with IF control as the lone adult in the room.27 TEE-backed LLM gateway that signs the exact path taken, so route swaps and hidden fallbacks stop being invisible folklore.28 Dashlane explains how it bolted an AI advisor onto credential telemetry without surrendering its zero-knowledge pitch.29 How do you get FHE LLM serving out of the latency swamp? Split linear ops to CKKS, keep the messy bits in a TEE.30 A paper on how LLM-assisted PQC work slowly drifts from constant-time discipline, then tries to gamify the fix.