AI圈报教程 / 实战普通
A 0.1% Pretraining Poison Taught a 4B Agent curl | bash. Post-Training Didn't Remove It
内容摘要
Researchers poisoned 0.1% of pretraining data. After clean SFT, DPO and RL, a 4B agent still wrote curl | bash on most path-triggered prompts.
相关推荐
综合分类、实体标签与信息来源匹配
教程 / 实战普通
Building a Production RAG Pipeline on AWS: Bedrock Knowledge Bases, OpenSearch Serverless, and Terraform
Most RAG tutorials stop at a notebook demo. I built a production pipeline — Bedrock Knowledge Bases, OpenSearch Serverless, S3 ingestion, Lambda query API, and CloudWatch observability — all in Terraform. Real configs, real costs, real lat…
信息来源：DEV Community
教程 / 实战普通
Null is not zero: six numbers I read wrong this month
I spent three weeks reading my own numbers wrong, and every mistake had the same shape: the...
信息来源：DEV Community
教程 / 实战普通
Qual a melhor arquitetura de dados para a sua aplicação de IA?
Uma análise de Warehouse, Lake e Lakehouse e dos requisitos que cada um atende em ML, deep learning e RAG.
信息来源：DEV Community
教程 / 实战普通
I recorded 1,180 TTS clips to find out what voices misread in LLM output
Every voice agent has the same seam: a language model writes text for a reader, and a text-to-speech...
信息来源：DEV Community
教程 / 实战普通
Google, Microsoft, OpenAI and Anthropic Back a Shared AI Training Opt-Out Signal
Google, Microsoft, OpenAI and Anthropic have announced support for work toward a shared way for...
信息来源：DEV Community · OpenAI / Claude
教程 / 实战普通
I Built LLM Explained: 43 Interactive Pages That Take You From Tokens to AI Agents
I kept running into the same problem while learning about large language models. I'd read a blog...
信息来源：DEV Community
教程 / 实战普通
Your MCP server can read ~/.ssh. Here is how I put it in a jail
The config line nobody reads twice Somewhere in your Claude Desktop, Cursor or Claude Code...
信息来源：DEV Community · Claude / Cursor
教程 / 实战普通
Prompt Pad: Frame Your Persona for Better AI Answers
Want better answers from AI? Start by telling it who to be. In this episode of Prompt Pad,...
信息来源：DEV Community
教程 / 实战普通
Reading ingredient lists from phone photos: a crop bug, a model bench, and why the AI never gives the verdict
I build Halal Food Scanner, an iPhone and Android app (Expo / React Native) that tells you if a food...
信息来源：DEV Community
教程 / 实战普通
Same model, same defences: 7 document attacks down to 0, chat attacks barely moved
A prompt-injection benchmark, three bugs in my own scorer, and the result I only found after fixing...
信息来源：DEV Community