Mythos goes official – and regulators respond differently

Highlight

Mythos goes official – and regulators respond differently

Ten days after the data leak, Anthropic has made it official: Claude Mythos Preview is here. The model is not only more powerful than anything previously available – within its first 48 hours, it identified thousands of zero-day vulnerabilities in critical software. Anthropic turned this into "Project Glasswing", an effort to deploy the model's offensive capabilities for defensive purposes. At the same time, the company is battling double outages and banning third-party frameworks from flat-rate subscriptions. The message: frontier AI is getting more powerful – and harder to control.

On the regulatory side, a striking contrast is emerging. While the EU is tightening binding rules for high-risk AI through the AI Act, the United Kingdom is taking the opposite path: the Treasury Committee is urging the Financial Conduct Authority (FCA) to act, but the FCA insists on its principles-based approach – no AI-specific rules, just existing frameworks. Two philosophies, one problem: how do you regulate a technology that grows faster than any regulatory framework can keep pace with?

Agentic AI

Tools, Skills & What's Trending

Anthropic News: Claude Mythos Preview – from leak to launch

Anthropic has officially unveiled Claude Mythos as a preview – ten days after a data leak revealed the model's existence. The remarkable part: within its first days, Mythos identified thousands of previously unknown security vulnerabilities in critical software. In parallel, Anthropic is launching "Project Glasswing" to channel these offensive capabilities into defensive cybersecurity. Also new: Managed Agents – hosted, scalable agent infrastructure for enterprise customers.

Open Source: Google Gemma 4 – agentic AI arrives on-device

Google DeepMind has released Gemma 4, a family of open-weight models that enable agentic workflows directly on-device – no cloud connection required. Four sizes (2B to 31B parameters), Apache 2.0 licence, 256K context window, and over 140 languages. The 31B model ranks #3 on the global open-model leaderboard. The strategic signal: agentic AI doesn't need to run in the cloud – and Google is providing the open-source foundation for it.

Market Trend: OpenAI – the $1 trillion IPO takes shape

OpenAI is preparing what could become the largest tech IPO (Initial Public Offering) in history. In its latest $122 billion funding round, $3 billion alone came from retail investors – a signal that CFO Sarah Friar reads as proof of broad trust in AI. Behind the scenes, however, tension is building: Friar is pumping the brakes on timing while CEO Sam Altman pushes for Q4 2026. With a projected annual loss of $14 billion, the question is whether markets will show patience for a company betting on growth – not profitability.

Banking & Regulation

Regulation & Financial Markets

UK Treasury Committee: AI in financial services needs clear guidance

The UK Treasury Committee has called on the FCA (Financial Conduct Authority) to publish practical AI guidance for financial services by the end of 2026. 75% of UK financial firms already use AI – yet the FCA insists on its principles-based approach without AI-specific rules. In parallel, Sheldon Mills is leading a long-term review of AI in retail financial services, with results expected in summer 2026. The contrast with the EU is stark: while Brussels is enacting detailed prescriptions through the AI Act, London is betting on flexibility – with the risk of acting too late.

DORA compliance: only half of institutions are ready

According to a Deloitte study, only 50% of financial institutions expected to be fully DORA-compliant (Digital Operational Resilience Act) by the January 2025 deadline – with a further 38% pushing their target into 2026. Meanwhile, BaFin published a new circular on 1 April, further specifying operational resilience requirements. The problem is structural: DORA, the AI Act and the MaRisk overhaul are all hitting institutions simultaneously, while compliance budgets cannot keep pace with the speed of regulation.

Signal & Noise

Worth Your Time

• Claude Mythos Finds Thousands of Zero-Day Flaws – The Hacker News
  Detailed technical analysis of how Mythos systematically identifies vulnerabilities in critical systems. Required reading for CISOs.
• Anthropic Blocks OpenClaw from Claude Subscriptions – TNW
  Anthropic bans third-party agent frameworks from flat-rate plans, forcing pay-as-you-go. A sign that the economics of autonomous agents remain unresolved.
• Accenture: The Era of Unconstrained Banking – Forbes
  Accenture's 2026 banking trends: AI is taking on real work, not just pilot projects. "Unconstrained Banking" describes an industry pushing operational boundaries through agents.
• Fintech 2026: Global Practice Guide – Chambers & Partners
  Convergence, agentic AI, regulatory legitimisation of digital assets and fraud liability shifts – the four megatrends according to Chambers.
• AI in Financial Services Risks Serious Harm – UK Parliament
  The full Treasury Committee report. Unflinching in its analysis, clear in its demand: the FCA must act before consumers are harmed.

"The first rule of any technology used in a business is that automation applied to an efficient operation will magnify the efficiency. The second is that automation applied to an inefficient operation will magnify the inefficiency."