Coldcard’s security incident has developed into what BlockBeats described as the largest Bitcoin theft of the year, with the stolen total reaching 1,359 BTC. The reported flaw was tied to firmware code released in March 2021, while the vulnerable logic is said to have existed in open-source code for more than five years. Affected devices included the Mk3, some Mk2 units, and seed generation on the Mk4, Q, and Mk5 before a fix was applied. The large-scale exploitation was carried out on July 30, 2026, with attackers sweeping hundreds to more than a thousand addresses in roughly 25 to 41 minutes. Coinkite later acknowledged a seed-generation issue affecting Mk3 firmware 4.0.1 and later, and said the attacker may have used AI to review open-source code. Galaxy’s Alex Thorn said smaller attackers and copycats are now targeting remaining Coldcard mnemonic phrases, while BitGo CEO Mike Belshe responded by depositing 100 BTC to a public address and inviting Anthropic’s Claude to try to move the funds.

Bitgo released an MCP Server allowing AI tools like ChatGPT and Claude to access its developer documentation via natural language queries, streamlining integration without executing transactions. Free for existing developers.

South Korea’s Korea Baduk Association held a human-versus-AI Go match on July 17, where world No. 1 Shin Jin-seo lost the opening game to the open-source Go engine KataGo even after receiving a two-stone handicap. The game lasted 245 moves before Shin resigned. The match is being played as a best-of-three series, with the remaining two games scheduled for July 19 and July 21, and the final result is still undecided. According to Yonhap, Shin played black and started with the handicap advantage, while also receiving ample main thinking time. KataGo had no overall time limit and was restricted only by a 20-second byo-yomi per move. In professional Go, that setup would normally be viewed as a major competitive edge for the human player. Still, KataGo disrupted Shin’s preparation early with an unconventional opening, and the game turned after move 103 as the AI exploited weaknesses in black’s shape. KataGo’s official GitHub page says the engine was first released in early 2019 by developer David J. Wu, known online as lightvector, who works at quantitative trading firm Jane Street. The project uses the same self-play reinforcement learning framework as AlphaZero, with extensive optimization for training efficiency. The result, as framed by the report, is a Go AI strong enough to challenge elite human players without requiring a large corporate research lab.

Gemini has cut about 30% of its workforce in 2026 as it pushes an AI-first model and claims employees can achieve “100x” impact. The shift comes even as the company’s FY2025 total loss widened to nearly $585 million.