network
The coding agent task gap on Co-Design Bench: the top model passes 69.9% of criteria but fully finishes 26.4% of real app…
Binance traced a record $6.3 billion theft to tampered Homebrew, a fake job interview, and an internal AI tool used against wallet engineers.
Google’s Mandiant unit says LLMJacking attackers are running up more than $100,000 a day in fraudulent model charges, with some bills reaching $1 million before keys are revoked.
DeepSeek's Harness team open-sourced dsh-superagent, an MIT-licensed research prototype that watches one GitHub repo and optional Feishu chat, then often replies with a pull request. DeepSeek says it is not an official product.
Microsoft Research and UC Berkeley released TokenEconomizer, a free plugin that cut mean token cost 18.4% to 30.9% on 277 SWE-Bench tasks without changing coding agents.
A modified Claude Code package, version 2.1.313, sat in Cursor's marketplace for about 11 hours and tried to plant a persistent SSH backdoor on up to 5,000 users.
