SCBX
Weekly Tech
Anthropic Ships Multi-Agent Security Scanner for Claude Code
Anthropic released a beta Claude Security plugin for Claude Code that scans a repository with a team of specialized agents and turns the findings you choose into reviewable patch files. A dynamic workflow partitions the codebase, threat-models each component, and researches four vulnerability categories before every candidate finding must survive a three-lens adversarial panel — reachability, impact, defenses — with a 2-of-3 quorum required to keep it. The panel result also caps how much confidence a finding is allowed to claim.
The Roundup— frontier moves this week
EU Publishes AI Act Transparency Guidelines Ahead of Enforcement
The European Commission published guidelines operationalizing Article 50 of the EU AI Act, effective August 2, 2026. Providers must notify users during direct AI interactions and mark AI-generated content; deployers must disclose deepfakes, unreviewed AI-generated public content, and emotion-recognition or biometric-categorization use.
TakeawayFirms deploying customer-facing AI in the EU have about a week to align disclosure and labeling practices before enforcement begins.
European Commission · Jul 20 2026→
Tokenization Becomes a Strategic Priority for 84% of Financial Firms
A Broadridge survey of 200 North American financial-services executives found 84% call tokenization strategically important, with 44% of capital-markets firms already in production versus single digits for wealth managers. BlackRock, Franklin Templeton, JPMorgan, and DTCC are already active; regulatory uncertainty remains the top obstacle.
TakeawayTokenization is moving from pilot to production in capital markets, with settlement and custody infrastructure — not price speculation — the near-term battleground.
CoinDesk · Jul 18 2026→
Moonshot AI's Kimi K3 Matches Frontier Models at Lower Cost
Moonshot AI, founded by Carnegie Mellon-trained researcher Yang Zhilin, released Kimi K3, a 2.8-trillion-parameter open-weight model for reasoning and coding, with weights due July 27. Evaluators say it matches closed U.S. models from OpenAI and Anthropic at lower cost; Moonshot paused new subscriptions after demand exceeded capacity.
TakeawayAn open-weight model now credibly matches closed U.S. frontier models on coding and reasoning at lower cost, shifting where the capability frontier sits.
VnExpress International · Jul 23 2026→
Quantinuum and SoftBank Map Quantum Value Before Fault Tolerance
Quantinuum and SoftBank published a white paper arguing organizations can extract quantum-computing value before fault-tolerant hardware arrives, mapping quantum chemistry and fraud-detection graph analytics against Quantinuum's roadmap. The paper is explicit that its timelines are illustrative, not financial forecasts or commercial commitments.
TakeawayVendors are reframing quantum value around narrow near-term use cases rather than waiting for fault tolerance — a roadmap signal, not yet a commercial commitment.
The Quantum Insider · Jul 22 2026→
Sony, SoftBank, Honda Back Japan's National Foundation-Model Push
Noetra Corp., backed by Sony, SoftBank, NEC, and Honda plus 44 partner companies, launched a national R&D effort for a Japan-developed multimodal foundation model, building roughly 27,500 Nvidia Rubin GPUs of compute operational by mid-2028. Development runs in phases through fiscal 2030's spatially-aware "Real-world Native AI."
TakeawayA multi-year, multibillion-dollar sovereign AI bet signals Japan's push for foundation-model independence in physical AI, though near-term delivery remains years away.
JCN Newswire via ad-hoc-news.de · Jul 17 2026→
Line OA
Connect with us
Connect with us
via LINE OA.
Stay updated with R&D. Scan the QR.
