---

In April 2026, as tech giants pour unprecedented billions into artificial intelligence infrastructure, a critical question looms over boardrooms and policy chambers alike: Are we building tools that will enhance human capability, or unleashing forces we’re unprepared to control? The recent revelation of Anthropic’s Claude Mythos model—deemed too powerful and dangerous for public release—represents a watershed moment that demands we confront an uncomfortable truth: the era of unregulated AI expansion may be ending, whether we’re ready or not.

The Mythos Moment

Claude Mythos is the first AI model Anthropic has publicly deemed too high-risk for public release. The Trump administration is trying to gauge the risks of Anthropic’s Mythos, an AI system it says can rapidly uncover — and exploit — digital vulnerabilities. This isn’t just another incremental advancement in AI capability—it’s a fundamental shift in how we must think about these systems.

Claude Mythos Preview is Anthropic’s most powerful AI model to date, and its cybersecurity implications are serious. Other frontier AI models—including OpenAI’s GPT-5.4-Cyber and Google’s Big Sleep—have some comparable capabilities already, and more will follow. The model was designed to push boundaries in software engineering, creating AI capable of working with vast, complex codebases in ways previous models could not. But that same capability makes it extraordinarily dangerous in the wrong hands.

The concern isn’t theoretical. A small group of unauthorized users have accessed Anthropic’s new Mythos AI model, a technology that the company says is so powerful it can enable dangerous cyberattacks. A handful of users in a private online forum gained access to Mythos on the same day that Anthropic first announced a plan to release the model to a limited number of companies for testing purposes. When your most powerful safety measure—restricted access—fails on day one, it underscores the near-impossibility of containing these capabilities once they exist.

The Cybersecurity Time Bomb

The implications for cybersecurity are staggering. The threat actor was able to use AI to perform 80-90% of the campaign, with human intervention required only sporadically. At the peak of its attack, the AI made thousands of requests, often multiple per second—an attack speed that would have been, for human hackers, simply impossible to match.

With the correct setup, threat actors can now use agentic AI systems for extended periods to do the work of entire teams of experienced hackers: analyzing target systems, producing exploit code, and scanning vast datasets of stolen information more efficiently than any human operator. Less experienced and resourced groups can now potentially perform large-scale attacks of this nature.

This represents a fundamental shift in the threat landscape. AI does not create new vulnerabilities, it exposes existing ones, making the chronic underinvestment that boards have tolerated for years an immediate and material business risk. Many organizations will need to significantly increase cybersecurity spending, by up to two times their current levels or even more; planned increases of about 10% annually fall far short of what the threat now demands.

Economic Disruption: The Workforce Reckoning

While cybersecurity threats grab headlines, AI’s impact on employment may prove even more socially destabilizing. The data paints a complex but concerning picture. Twenty-seven percent of employees in AI-adopting organizations say that their workplace has changed in disruptive ways to a large or very large extent in the past year. By comparison, 17% of employees in organizations that have not adopted AI report the same level of disruption.

Employment growth in industries such as marketing consulting, graphic design, office administration, and telephone call centers has fallen below trend amid reports of reduced labor demand due to AI-related efficiency gains. In addition, employment growth in technology-sector occupations such as computer systems design, software publishing, and web search portals has slowed sharply.

The impact on younger workers is particularly stark. Unemployment among 20- to 30-year-olds in tech-exposed occupations has risen by almost 3 percentage points since the start of 2025, notably higher than for their same-aged counterparts in other trades and for overall tech workers as well. Stanford’s Canaries in the Coal Mine 2025 report revealed the U.S. jobs most exposed to AI automation have declined fastest for ages 22-25.

Perhaps most troubling, AI is behind at least some layoffs, but these are almost completely in anticipation of AI’s impact—companies are cutting workers based on what AI might do, not what it’s currently doing. Many enterprises, despite how ready or not they are to successfully use AI solutions, will say that they are increasing their investments in AI to explain why they are cutting back spending in other areas or trimming workforces.

The Governance Gap

The most alarming aspect of current AI development isn’t the technology itself—it’s the yawning gap between capability and control. The real risk in technology right now is that companies have moved so fast that few built the enterprise architecture to hold it together. That absence of enterprise governance is costing organizations in ways the adoption numbers do not show.

As organizations continue racing to adopt and scale generative AI, 2026 is shaping up to be a year defined by both acceleration, correction, and control. The pace of innovation remains staggering while governance, evaluation, and risk management have struggled to keep pace.

Even leading AI companies recognize they’re pushing into uncharted territory. While we do not believe Claude Opus 4.6 meets the threshold for ASL-4 autonomy safeguards, we find ourselves in a gray zone where clean rule-out is difficult and the margin to the threshold is unclear. When the developers themselves can’t clearly assess whether their own safety thresholds have been crossed, we have a problem.

The Regulatory Response

Governments are scrambling to respond. The AI Act entered into force on 1 August 2024, and will be fully applicable 2 years later on 2 August 2026, with some exceptions. By the end of 2026, “death by AI” legal claims will exceed 2,000 due to insufficient AI risk guardrails. Black box systems — AI models whose decision-making processes are opaque or difficult to interpret — can misfire, especially in high-stakes sectors like healthcare, finance and public safety.

At the state level in the US, regulations are proliferating. Several states have enacted or finalized broad AI governance statutes that impose affirmative risk management, documentation, and oversight obligations for certain high-impact AI systems, with enforcement beginning in late 2025 and 2026. However, The U.S. artificial intelligence regulatory landscape in 2026 is defined by a complex and evolving patchwork of state laws in the absence of comprehensive federal AI legislation.

The Infrastructure Arms Race

Meanwhile, the financial stakes continue to escalate. Google plans to invest up to $40 billion in Anthropic. The pact involves an initial $10 billion from Google at Anthropic’s latest valuation of $380 billion. Anthropic’s annualized revenue has topped $30 billion.

The AI race is increasingly defined by access to the compute needed to train and deploy these systems. This creates perverse incentives—companies must secure massive infrastructure commitments before fully understanding the risks of what they’re building.

Why Careful Management Matters

The case for carefully managed AI development isn’t about slowing innovation—it’s about ensuring innovation doesn’t outpace our ability to safely deploy it. Several principles should guide this approach:

First, transparency must be non-negotiable. AI technologies need to be safe and transparent. There are few, if any, benefits from being outside efforts to achieve this. When companies like Anthropic restrict their most powerful models, they’re acknowledging that not every capability should be immediately commercialized.

Second, governance must precede deployment. With agentic systems comes a new class of risk, and 2026 will be the year organizations begin addressing it seriously. Expect advances in areas like guardrails, human-in-the-loop feedback, audit trails, and end-to-end testing for agents.

Third, society needs time to adapt. By 2030, AI could automate 30–70% of tasks in many sectors, but human-AI collaboration will create new opportunities if societies adapt via reskilling and policy. The “if” in that statement is doing heavy lifting—adaptation requires investment, planning, and time that pure market forces don’t naturally provide.

Fourth, cybersecurity fundamentals must be strengthened immediately. The immediate priority is strengthening cybersecurity fundamentals: Strong foundations provide significant protection against AI-enabled attacks, and most organizations urgently need to build those foundations.

The Path Forward

Many countries are rightly being cautious and assessing risks, but more coherence is needed in policymaking. Nations should work together to design policies that not only enable development, but also incorporate guardrails.

The Anthropic Mythos moment should serve as a catalyst for honest conversation about AI development trajectories. We’re not facing a binary choice between innovation and safety—we’re facing a choice between thoughtful, managed development and a headlong rush that may produce capabilities we’re unprepared to handle.

If 2025 has been the year of AI hype, 2026 might be the year of AI reckoning. Its powerful capabilities are already driving advances in healthcare, manufacturing and more, yet, in some areas, the returns on investment are mixed and potential future profits not certain.

The question isn’t whether AI will transform society—it already is. The question is whether that transformation will be deliberate and equitable, or chaotic and destabilizing. Models like Claude Mythos demonstrate both AI’s extraordinary promise and its genuine dangers. The decisions we make now about how to manage this technology’s growth will echo for generations. We cannot afford to get this wrong.

Amit Shrivastava