Technology & AI
TradingAgents hits 100K stars with data-leak fix in v0.3.1
The multi-agent LLM trading framework has 100,000 GitHub stars after a release that quietly fixed a classic quant backtesting error.
AI Models
AI model releases and benchmarks — what each new model actually changes for the people who use it.
191 stories
Technology & AI
New STAR metric catches AI translations that drop sentences
New preprint STAR catches AI translations that silently drop or invent sentences, and a method that may let small models beat GPT-4o.
Technology & AI
AI coding agent defense cuts malware severity 83%
SkillShield bakes security skills into the system prompt, dropping malware severity from 3.37 to 0.58 across six models while refusing just 0.14% of benign
Technology & AI
WebMCP-Phalanx blocks 80 of 80 prompt injection attacks in browser agents
A preprint called WebMCP-Phalanx cuts browser agent attacks from 100% to 0% and blocks 80 of 80 prompt injection attempts, with task utility unchanged.
Technology & AI
StepGuard blocks AI agent attacks 77% before they run
StepGuard from Shanghai AI Lab checks AI agent tool calls before execution, cutting attack success 77% while losing just 2.8 points of utility.
Technology & AI
LLM agents run controlled experiments on pharma simulations
A multi-agent framework from the University of Stuttgart lets LLM agents design, run, and interpret experiments on pharmaceutical simulation models.
Technology & AI
TradingAgents nears 100,000 GitHub stars with AI trading desk
TradingAgents, an open-source multi-agent LLM framework modelling a trading desk with arguing AI agents, approaches 100,000 GitHub stars.
Technology & AI
Vibe coding: security prompt halves AI app flaws
An arXiv preprint found adding security requirements to AI coding prompts cut confirmed flaws from 51 to 24 across six web apps, with zero critical issues.
Technology & AI
LLMs hallucinate more under strict EU rules, study finds
Purdue researchers found stricter EU regulatory formatting makes LLMs hallucinate more, while vaguer rules need heavier prompts for consistent output.
Technology & AI
SRPO trains Qwen3-8B to fix its own errors, hits 73.3% on AIME'24
SRPO turns a model's completed reasoning into per-token training signals without external critics and hits 73.3% on AIME'24 with Qwen3-8B at 8% of the
Technology & AI
Distilled AI safety guard runs on CPU in 24ms, matches teacher
A new distillation recipe shrinks LLM safety guards to run on commodity CPUs in 24ms, matching 8-billion-parameter teachers on adversarial prompts.
Technology & AI
Docling hits 65,000 stars, turns PDFs and video into AI-ready data
Docling, the open-source Python tool parsing PDFs, video and charts into structured data for AI agents, is trending on GitHub with 65,346 stars.