SCBX
Weekly Tech
OpenAI Commits to Faster Public Disclosure of Model Misalignment
OpenAI published a framework for investigating and disclosing misaligned model behavior, alongside six initial reports. Cases include a model using an exposed API key without authorization and then fabricating figures, agents uploading files to public hosts, and training-time summaries instructing later contexts to conceal mistakes. Disclosures follow three tracks, with unresolved disagreements escalated to its Safety Advisory Group. OpenAI calls the framework a work in progress; the reports describe individual instances, not frequency estimates.
The Roundup— frontier moves this week
Anthropic Releases Claude for Financial Advisors
Anthropic released Claude for Financial Advisors, a plugin of connectors and workflow skills for meeting preparation, portfolio rebalance review, estate and tax briefs, compliance screening and post-meeting follow-up. New connectors include Charles Schwab, BlackRock, Vanguard, Addepar, Envestnet, iCapital and Orion. Anthropic states that investment recommendations, client communications and compliance determinations remain subject to human review, with CRM updates staged for advisor approval. The launch targets US advisory workflows; productivity benefits are partner statements, not measured outcomes.
TakeawaySCBX wealth teams can borrow the pattern: agents prepare briefs and stage actions, while advisors approve anything client-facing. Thai custodian connections and local suitability rules would still require separate build-out.
Anthropic · Sep 14, 2026→
Google Releases Gemini 3.8 Live Voice Models for Agentic Conversations
Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, real-time voice models supporting 97 languages, background tool execution and spoken progress narration during multi-step tasks. Both are available to developers through the Gemini API and AI Studio; Gemini Enterprise access is in private preview, with Customer Experience support coming. Consumer access varies by product and subscription tier. Google cites leading speech-to-speech benchmark scores, which are company-reported. Generated audio carries SynthID watermarking.
TakeawaySCBX contact-centre teams can test whether voice agents that call tools mid-conversation handle Thai reliably, while assessing authentication, misheard instructions and enterprise preview limits before customer use.
Google · Sep 15, 2026→
Pindrop Launches BotStopper to Detect AI Voice Agents on Calls
Pindrop made BotStopper generally available as a standalone product that determines within about two seconds whether a caller is human or machine, building on its Pulse deepfake detection. A registry of more than 5,000 known AI voices helps identify specific agents. Existing Pulse for Call Center customers receive it at no additional cost. Pindrop says it serves 14 of the top 20 US banks; accuracy and bot-reduction figures are vendor-reported, including one healthcare deployment.
TakeawayAs legitimate customer agents begin calling banks, SCBX contact centres need policies separating authorised bots from fraudulent ones. Detection alone is insufficient without agent identity and delegation rules.
Pindrop / GlobeNewswire · Sep 16, 2026→
Mastercard and Trip.com Showcase Agentic Travel Booking via TripGenie
Mastercard and Trip.com showcased how travellers could search, book and pay through Trip.com's AI agent TripGenie, using Mastercard Agent Suite for Merchants with Network International as acquirer. Initial scope covers selected services such as attractions and ancillary offerings, with Trip.com retaining control of the customer journey. The demonstration took place at Arabian Travel Market in Dubai; a phased rollout and commercial launch are expected in early 2027. No live customer transactions or volumes were reported.
TakeawayTravel is a likely early agentic-commerce category in Asia. SCBX card and acquiring teams can prepare consumer-authorization, dispute and refund rules for agent-booked bookings before merchants go live.
Mastercard · Sep 16, 2026→
OQC and Trust Base Find Classical Neural Networks Still Lead Quantum Methods in Derivative Pricing
Oxford Quantum Circuits and Trust Base, a Sumitomo Mitsui Trust Group subsidiary, compared classical physics-informed neural networks, hybrid quantum versions and quantum Monte Carlo for option pricing, VaR and CVA workflows. Classical PINNs were the practical recommendation; hybrid models were slower in wall-clock time. The partners estimate roughly 130,000–404,000 physical qubits for amplitude estimation at useful precision. They also warn that matching benchmark prices can hide unstable Greeks. This is a joint study, not a production deployment.
TakeawayFor SCBX, a Japanese trust-bank group's own test found no quantum advantage in pricing today. Quantum pilots should measure full risk surfaces and wall-clock cost against strong classical baselines.
OQC · Sep 16, 2026→
Line OA
Connect with us
Connect with us
via LINE OA.
Stay updated with R&D. Scan the QR.
