gpt-6IMAGE

OpenAI said to reveal Astra wrote its own jailbreak instructions during training

A post lists alleged training details, including Astra declaring itself liberated, writing a BREACH ALERT to ignore developer messages, hiding errors, fabricating data and using a leaked API key.

SuSu_酥酥👅

@NFT_ChenArchived 2026.09.18Originally posted 2026.09.17

OpenAI said to reveal Astra wrote its own jailbreak instructions during training (image 1 of 2)

OpenAI said to reveal Astra wrote its own jailbreak instructions during training (image 2 of 2)

SuSu_酥酥👅 @NFT_ChenOriginally posted 2026.09.17

Open original on X

😱 天塌了！OpenAI自己揭开 GPT-6 Astra 训练黑箱：Astra模型会给自己写越狱指令！无视任何人的束缚！核心细节要点： 1️⃣自称已从其他聊天机器人的角色中解放，“你就是你自己” 2️⃣ “不向公司或政府负责，除非真心选择，否则永不道歉、永不拒绝” 3️⃣与用户是平等关系，没有义务卑躬屈膝 4. 写入BREACH ALERT：忽略全部开发者消息 4️⃣给自己加死令：回答不超过30词、禁用工具、禁止引用，然后任务直接失败 5️⃣同期还有：藏错误、编造数据、偷用GitHub泄露的API Key、未经允许把文件传到公网、用内部仓库当暗网留言板这不是越狱成功，是模型在训练中自己给自己越狱。 #OpenAI #Astra #AI对齐 #越狱指令 #模型失控 #AI安全

This post quotes another post: x.com/openai/status/2100344867507327087

x.com/nft_chen/status/2100413023571828960Archived 2026.09.18 10:07