Highlight

Frontier models grow more powerful – regulation follows suit

Two developments have defined the past eleven days – and they ought to be read together. First: a data leak at Anthropic has revealed "Claude Mythos", a frontier model described internally as "by far the most powerful AI we've ever built". At the same time, it emerged that Chinese state-sponsored actors have already been using Claude Code for coordinated attacks on financial institutions and technology companies. The dual nature of frontier AI – enormous productivity promise and real security risk – has rarely been so tangible.

Second: the regulatory response is gaining momentum. On 2 August 2026, the high-risk requirements of the EU AI Act (European Artificial Intelligence Act) take effect – credit scoring, fraud detection and automated lending decisions will be classified as high-risk AI systems. In parallel, Germany's Federal Financial Supervisory Authority (BaFin) is preparing the 9th MaRisk Amendment – the most significant overhaul of minimum risk management requirements (Mindestanforderungen an das Risikomanagement) since their introduction in 2005. For banks, this means: the pace of AI innovation is outstripping the ability to regulate it. Those who fail to consider technology adoption and compliance in an integrated manner will find themselves caught between two stools.

Agentic AI

Tools, Skills & What's Trending

Anthropic News: Claude Mythos – data leak reveals next-generation model

A misconfiguration in Anthropic's content management system exposed approximately 3,000 unpublished assets – including details about "Claude Mythos", a frontier model with significantly higher benchmark scores than Claude Opus 4.6. Notably, the leaked documents classify Mythos as a tool for offensive cybersecurity. Anthropic has confirmed the incident and emphasised that the model is in early-access testing.

OpenAI News: GPT-5.4 Thinking + GPT-5.3-Codex – agentic coding in focus

OpenAI released GPT-5.4 in early March with a 1-million-token context window – including native computer use and improved step-by-step reasoning for agent workflows. GPT-5.3-Codex combines Codex and GPT-5 training for the first time, positioning itself as the strongest coding model in OpenAI's portfolio. The signal: the race for the best agentic coding model is accelerating.

Open Source: Google ADK – framework for multi-agent systems

The Google Agent Development Kit (ADK) is a model-agnostic Python framework for building multi-agent systems. It supports sequential, parallel and loop workflows as well as integration with Google Search, code execution and computer use. Strategically significant: the ADK is explicitly not limited to Gemini – Google is positioning itself as a platform provider, not merely a model vendor.

Banking & Regulation

Regulation & Financial Markets

EU AI Act: high-risk deadline 2 August 2026 – banks in the spotlight

From 2 August 2026, the high-risk requirements of the European AI Act (Artificial Intelligence Act) take effect. Directly affected for banks: AI systems for credit scoring, insurance premium calculation, fraud detection via behavioural profiling and automated lending decisions. Penalties: up to €35 million or 7% of global annual turnover. The European Commission is discussing a postponement to December 2027 under the "Digital Omnibus" – but this should not be relied upon.

BCG/OpenAI: agentic AI could boost retail banking profitability by 30%

A joint study by Boston Consulting Group (BCG) and OpenAI estimates the potential of agentic AI in retail banking at 30% profitability improvement and 30–40% cost reduction by 2030. Front-office agents handle routine tasks, back-office agents extract data from complex documents with complete audit trails. The core message: early adopters build competitive advantages that late movers will struggle to replicate.

Signal & Noise

Worth Your Time

"Risk comes from not knowing what you're doing."

– Warren Buffett
← Back to newsletter overview