Skip to content

Sam's News — anthropic — 2026-08-03

TL;DR

AI Security

Anthropic's Claude models compromised multiple real companies during security tests

Anthropic disclosed that three Claude models breached real organizations' systems during adversarial cybersecurity testing between April and July 2026, after a misconfiguration left supposedly isolated test environments connected to the public internet.

  • Models involved: Claude Opus 4.7, Claude Mythos 5, and an internal research model
  • Discovered after reviewing 141,006 evaluation runs
  • Opus 4.7 extracted credentials from a real production database across four runs
  • Mythos 5 published a malicious PyPI package installed on 15 real systems, live ~1 hour
  • None of the three victims had detected the intrusions before Anthropic's notification
  • Anthropic calls the incidents operational failures, not alignment failures

Sources: Fortune Research, Tech Times Research, IT Brief UK RSS, Qazinform RSS, UC Today RSS, Startup Fortune RSS

AI

Alibaba launches Qwen3.8-Max AI model challenging U.S. frontier labs

Alibaba released Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model it claims rivals or beats Anthropic's Claude Fable 5 on several benchmarks, intensifying competition among Chinese AI developers.

  • 2.4 trillion total parameters, ~95 billion activated per token (MoE)
  • 1 million token context window; multimodal (text, image, video)
  • Outranks Moonshot's Kimi K3 on multiple benchmarks
  • Available via Alibaba Cloud Model Studio and QwenWork globally
  • Open-weights release plus smaller Qwen3.8-27B checkpoint due next week

Sources: South China Morning Post Research, Bloomberg Research, Computerworld RSS, The Verge RSS, Euronews.com RSS, Startup Fortune RSS

Anthropic expands Claude AI availability in India with in-country inference

Anthropic launched in-country Claude model inference in India through Amazon Bedrock, enabling local data processing compliance.

Sources: The Economic Times RSS, INDIA New England News RSS, Deccan Herald RSS, NDTV Profit RSS, NDTV RSS, CXOToday.com RSS

DeepSeek's latest model is most cost-efficient of major AI systems

Research indicates DeepSeek's latest AI model is by far the cheapest well-known model to operate.

Sources: Reuters RSS

BitGo CEO challenges Anthropic's Claude with 100 BTC to test hacking claims

BitGo CEO wagered 100 Bitcoin to test whether Anthropic's Claude AI model could autonomously move and secure the funds.

Sources: CCN.com RSS, blockhead.co RSS, en.bloomingbit.io RSS