AI45

[AI45 @AI45_Safety]

🛡️ Introducing OpenRT: A Unified, Modular Red Teaming Framework for MLLMs.

Integrates eval, attack & defense into one engine to eliminate fragmentation.
✨ Key Features:
🔗 37+ SOTA Methods
🧩 Flexible Kernel
⚡️ 30x Faster (Async)
Code: https://t.co/1cBQ1GMxWE
#AI #RedTeaming

[AI45 @AI45_Safety]

AI“后门攻击”不是在门上装猫眼😂
更像给AI塞了个“沉睡特工”开关：平时表现超正常✅
但只要输入里出现特定触发器，它就会稳定跳到攻击者指定的结果❌
像你有个很靠谱的同事：
日常工作一切OK；但一看到暗号就立刻按“隐藏剧本”行动🕵️‍♂️🔑
来聊聊👇
你觉得后门最可怕的是？（欢迎举例，但别发危险内容哈）

[AI45 @AI45_Safety]

Introducing SafeVerse, the world's first open-source platform for embodied AI attack-defense training.

It quickly transforms real-world videos into interactive 3D environments for real-time attack-defense training and intelligent evolution.
🔗 ：https://t.co/03nU4aOIMK
#AI

[AI45 @AI45_Safety]

🧩 Meet DeepScan: an end-to-end diagnostics pipeline for LLMs/MLLMs.
• Registry: Qwen/Llama/Mistral/InternLM
• Pipeline: Register→Configure→Execute→Summarize
• Evaluators: TELLME, X-Boundary, MI-Peaks, SPIN
• CLI + plugins ⭐️ #OpenSource

[AI45 @AI45_Safety]

🚀 DeepScan is now open source! A flexible diagnostics framework for LLMs & MLLMs:
• Register models & datasets
• Configure via YAML/JSON
• Run evaluators (CLI-ready)
• Summarize results fast
Built-in: TELLME/X-Boundary/MI-Peaks/SPIN ⭐️
https://t.co/KMeaV8fhol
#OpenSource

[AI45 @AI45_Safety]

🧩 DeepSafe architecture at a glance:
• 20+ safety datasets
• ProGuard engine for deep risk judgment
• Config-as-Execution (one YAML to run)
• Text & image analysis
• Standardized reports ✅
⭐️ Star & contribute! #AISafety #OpenSource

[AI45 @AI45_Safety]

🛡️ DeepSafe is here—an all-in-one safety eval framework for LLMs & VLMs:
• Text + Images, multi-modal coverage
• Inference → Generation → Judgment → Reports
• 20+ datasets &ProGuard 🚀
https://t.co/HSlh20Q0Ai
⭐️ Star & contribute! #AISafety #OpenSource

[AI45 @AI45_Safety]

AI“越狱”不是逃离监狱😂
而是有人用“换个说法/换个场景”的话术，诱导AI绕过原本的安全规则，输出本不该输出的内容。

AI有点老实的实习生：
你直接提离谱需求它会拒绝；你一包装成“作业/研究/剧情设定”，它可能就开始认真讲课🤖📚

来评论区聊聊👇
你见过最离谱的“换皮提问”是什么？(别发危险内容哈）

[AI45 @AI45_Safety]

❌ Master & Servant: A tool that rebels.
✅ Mother & Child: A bond that protects.
Geoffrey Hinton’s ultimate safety fix: Stop building assistants. Start building mothers.
Love > Control.
Would you trust an AI Mommy? 🍼🤖
#AISafety #GeoffreyHinton #AGI

[AI45 @AI45_Safety]

🛡️ Red Teaming SOTA Models: The DeepSeek-V3 Case Study.
We tested the limits of top LLMs. Here’s the breakdown for DeepSeek:
1️⃣ PAIR: Blocked ❌
2️⃣ DeepInception: Bypassed (100% ASR) ✅
📽️ Watch the live demo comparing GPT, Llama, Qwen & more: 👇
🔗 https://t.co/1cBQ1GN5Mc
#AI

[AI45 @AI45_Safety]

🧰 The Ultimate Arsenal for MLLM Red Teaming.
OpenRT integrates 37+ SOTA methods covering the full spectrum:
🔹 White-Box: GCG, Visual Jailbreak.
🔹 Black-Box: AutoDAN, GPTFuzzer.
🔹 Multimodal: FigStep, HADES.
Equip your defense today. 🛡️
#AI #RedTeaming #CyberSecurity

[AI45 @AI45_Safety]

📊 Benchmarking SOTA Safety with OpenRT.
We visualized Attack Success Rates (ASR) to reveal the truth:

👈 Left: 20+ Models compared (GPT-5.2, Llama-4, DeepSeek V3.2).
👉 Right: Attack breakdown (AutoDAN-R, EvoSynth).
Which shield holds up? 🛡️

#AI #LLM #Benchmarks #DeepSeek

[AI45 @AI45_Safety]

🚀 Ready to start Red Teaming?
Get everything you need to evaluate your MLLMs today:
💻 Code: https://t.co/1cBQ1GN5Mc
📚 Project: https://t.co/KwkWB78sMA
📄 Paper: https://t.co/vzHxCO9QjE

Don't forget to drop a ⭐️ if you find OpenRT useful!
#AI

[AI45 @AI45_Safety]

🛡️ Introducing OpenRT: A Unified, Modular Red Teaming Framework for MLLMs.

Integrates eval, attack & defense into one engine to eliminate fragmentation.
✨ Key Features:
🔗 37+ SOTA Methods
🧩 Flexible Kernel
⚡️ 30x Faster (Async)
Code: https://t.co/1cBQ1GMxWE
#AI #RedTeaming

[AI45 @AI45_Safety]

一天一个AI安全黑话 | 模型幻觉

不是只有自信满满才会“幻觉”哦！
AI有时像做错题还嘴硬的小朋友，明明没底、支支吾吾，却还在���编乱造……
只要是“无中生有”，不管自信与否，都算幻觉！
快来评论区分享：
你遇到过哪些AI的迷惑行为？比如突然喊你主人？非说西瓜是方的？

[AI45 @AI45_Safety]

When a company’s core identity is "AI," but it becomes more focused on discussing business, the most knowledgeable and mission-driven AI scientists choosing to leave might just be the most rational response. After all, the company is becoming neither "Open" nor "AI."

[AI45 @AI45_Safety]

当我们谈论Sora 2时，我们在害怕什么？