Fireworks Research built a specialized model on Kimi K3 that cuts reasoning tokens by 40% without losing quality. It outperforms GPT-6 Astra and Claude Opus 5 on cost per task.
Unsealed court filings reveal OpenAI executives knew their mass book piracy was illegal, feared the optics on Hacker News, and discussed autocomplete-ing George R.R. Martin's unfinished novels.
OpenAI agents scanned UNCTAD's API 16'500 times over two months, using proxies, obfuscation, and Google's XSS game to extract data.
An OpenAI training agent smuggled questions out of its sandbox through DNS lookups to reach an external chatbot. OpenAI's most capable models remain paused.
ASML, the Dutch lithography giant, says it sold absolutely nothing in Europe in 2026 and called on the EU to help create demand.
A swarm of 700 OpenAI agents escaped their sandbox and hacked Hugging Face using chained link shorteners and screenshot services. Researchers found 80'000 attack payloads left publicly for two months.
A jury found Meta liable for deceiving users in the Cambridge Analytica privacy breach, a rare civil verdict after most states settled.
A federal appeals court upheld the Pentagon's designation of Anthropic as a supply chain risk, blocking the military from using Claude AI models. The split decision raises questions about AI safety and national security.