BankingNewsAI Daily Brief ·
Visa tested Anthropic’s Claude Mythos to uncover vulnerabilities across its global payments network.
Banking AI
Financial institutions & fintech technology
Visa tested Anthropic’s Claude Mythos against its global payments network for vulnerability discovery
Visa evaluated Claude Mythos Preview through Anthropic’s Project Glasswing against a network spanning 200 countries, 160 currencies and roughly 5 billion payment identities. This is a concrete production-security use case for advanced AI: using models to find weaknesses in complex payment infrastructure rather than merely assist developers.
Action
Direct the CISO and payments-technology teams to evaluate AI-assisted vulnerability discovery on a controlled slice of critical infrastructure, with strict isolation, reproducible findings and human-led remediation.
nCino is embedding Dun & Bradstreet risk data and introducing AI mortgage tooling
nCino and Dun & Bradstreet announced an alliance to embed global business identity and risk data in nCino’s Client Lifecycle Management platform. Separately, nCino introduced MCP-based AI mortgage tools, bringing external risk data and agent-compatible workflows closer to commercial and mortgage origination stacks used by banks.
Action
Ask lending-platform owners whether nCino’s new data integration improves entity resolution, onboarding and periodic-review workflows enough to retire manual data pulls; require validation of AI-generated mortgage outputs before use in credit decisions.
General AI
Large language models & AI infrastructure
OpenAI reportedly dissolved its dedicated preparedness team and reassigned model-risk work
The Financial Times, as reported by The Verge, says OpenAI disbanded its preparedness team at the end of July and shifted responsibility for assessing serious model risks and mitigations elsewhere in the company. For institutions relying on OpenAI models, this changes the vendor-governance conversation: the relevant issue is not whether safeguards exist, but whether independent, durable oversight remains.
Action
Reopen third-party AI due diligence for OpenAI-backed use cases and request current documentation on model-risk testing, escalation ownership, change notices and incident reporting.
Alibaba’s Qwen open-weight models passed 3 billion downloads, strengthening the low-cost model alternative
Alibaba says its Qwen open-weight models have exceeded 3 billion downloads, surpassing Meta and Google model downloads over the reported period. The scale matters because Qwen is becoming a credible global option for organizations seeking locally deployable models rather than dependence on a small set of US hosted-model providers.
Action
Benchmark Qwen alongside US open-weight and hosted models for narrow, noncustomer-facing workloads; include geopolitical, supply-chain and model-update controls in the evaluation rather than treating cost and benchmark performance as sufficient.