proomt

Search

Search posts, papers, and topics

All posts

Hugging Face Daily PapersZhuoXin Liu, Zhiming Ma, Ying Zhang1 min readpaperadvanced

RiskChainBench: A Benchmark for Obfuscated Platform Message Restoration and Evidence-Grounded Web Investigation

Read on arxiv.org

Summary

RiskChainBench is a new benchmark that pairs synthetic obfuscated message restoration inputs with human‑labeled local web environments, requiring models to both decode malicious instructions and investigate the linked site. Across ten models, restoration accuracy varies widely and web‑agent failures dominate the error budget.

  • The benchmark provides 3,600 synthetic restoration examples linked to 600 human‑annotated sandboxed web sessions.
  • Evaluation is split: Task 1 restores message intent and destination; Task 2 runs a VLM‑driven web agent to produce an evidence‑cited risk report.
  • Top‑1 entry recovery ranges from 35.2% to 95.2%; downstream web decision accuracy is only 26.3%–62.8%, highlighting a large gap.
  • 31.9% of web runs fail due to exploration errors, while post‑decision typing errors are under 1%, pinpointing exploration as the main bottleneck.

Anyone building multimodal LLM agents for detecting malicious content needs an end‑to‑end benchmark that measures both decoding of obfuscated instructions and safe web investigation.

7/10
llmmachine learningsecurityml infra

Related reading

  2. We now auto-approve and merge 15% of PRs

    This article describes how a team implemented a bot to auto-approve and merge 15% of low-risk pull requests, aiming to reduce code review bottlenecks and improve development velocity. The system relies on custom deterministic linters, basic end-to-end tests, and a risk scoring mechanism to identify safe changes.

    Swizec Tellerswizec.com3 min

  3. Poisoned Documents, Real Risks: Sebastián Passaro Puts AI’s Weakest Link to the Test at Testear.la 2026

    Sebastián Passaro (Qubika) demonstrated a live RAG pipeline attack at Testear.la 2026, showing how a single poisoned document can hijack LLM outputs and trigger unsafe actions. He tied the demo to the OWASP LLM Top 10, highlighted open‑source tooling for finding such weaknesses, and advocated a defense‑in‑depth threat model for QA teams. The talk reframed AI from a testing aid to a security surfa…

    Moove-itqubika.com4 min

  4. GPT-6 Astra Is the First Model OpenAI Classifies as Critical for Cybersecurity

    OpenAI labeled GPT‑6 Astra as “Critical” for cybersecurity under its Preparedness Framework – the first model to meet that bar. In controlled tests the model autonomously discovered zero‑day bugs in a browser and an OS kernel, building working exploit chains in 29 h (browser) and 12 h (kernel). A benchmark of post‑cutoff vulnerabilities confirmed its ability to find unknown flaws. OpenAI reports…

    InfoQinfoq.com3 min

  5. FRAUDSkill: Structured Frozen-Weight Skill Optimization for Audio Anti-Fraud Detection

    FRAUDSkill is a framework that keeps a pretrained audio‑language model frozen and learns an external layer of skill programs, routing policies, and decision rules to meet a structured anti‑fraud detection protocol. On the TeleAntiFraud benchmark it reaches 73.5% Macro‑F1 (≈32% improvement) while cutting invalid predictions to 1.94%.

    Hugging Face Daily Papersarxiv.org1 minpaper

  6. Grounded Skill Synthesis from Code at Scale for Agentic Intelligence

    Code2Skill is an automated pipeline that mines popular GitHub repositories to extract verifiable, implementation‑anchored procedural “skills”. It builds a bank of ~1 M skill records (atomic ops, workflows, patterns) with provenance metadata, verifies each via blind reconstruction, and shows that augmenting LLM‑based agents with these skills yields an average 11.7% performance lift across 72 proto…

    Hugging Face Daily Papersarxiv.org1 minpaper