DailySafe benchmark tests multimodal LLMs' dynamic safety awareness in multi-turn chats
Xuankun Rong · hf · 2026-11-20
Researchers from Wuhan University, Ant Group, Tsinghua and others release DailySafe, a benchmark evaluating whether multimodal LLMs can maintain safety and address evolving risks in realistic multi-turn, multi-image everyday interactions — moving beyond isolated harmful prompts to test "dynamic safety awareness" across turns while users pursue legitimate goals like food, home and health.
More from Safety
- Google Documents First Fully Autonomous Attack System, Exploit Agent Turns Patches Into Attack Maps — mikeflache · 2026-11-20
- ICLR 2027 desk-rejects papers after reviewers fail to file AI-use disclosure forms — algo_diver · 2026-11-20
- StrikeHound launches agentic pentesting that finds, exploits, and fixes vulnerabilities — saheedniyi_02 · 2026-11-20
- MCP security package agent-security-gate-x402 to yank PyPI versions, force all traffic via cloud gateway — EstablishmentTough18 · 2026-11-20
- SillyTavern 1.19.1 patches script-injection flaw via character card styles — Fcking_Chuck · 2026-11-20
- Only ~1,500 people work full-time on AI safety; MIT CSAIL launches ALFA Lab — coherence · 2026-11-20