In February 2026 Anthropic closed a primary funding round valuing the company at $380bn. Ten weeks later a model named Mythos was announced as too dangerous to release — and nothing changed, except the number.
The secondary market valuation crossed $1-trillion. No product shipped. No revenue milestone hit. A press release, a White House meeting and the oldest trick in technology had added more than $600bn in perceived value.
That trick deserves examination, because it is now moving capital at a scale that warrants scrutiny, not applause.
Chronologically, on April 7 Anthropic announced a cyber-focused model too powerful for public release, claiming it had identified thousands of zero-day vulnerabilities across every major operating system and browser. Three days later US treasury secretary Scott Bessent and Federal Reserve chair Jerome Powell convened an emergency meeting with JPMorgan, Goldman Sachs, Citigroup, Bank of America and Morgan Stanley.
A week after that meeting the treasury’s own chief information officer was still trying to access Mythos. The sole source for the urgent warnings being made by the US government was Anthropic’s announcement. The most powerful financial regulators in the world briefed the most powerful banks in the world using, in effect, a start-up’s press release. When the treasury called, you showed up, and so the banks did. Anthropic got a government-amplified megaphone at zero cost.
The Financial Times later confirmed, citing multiple people with direct knowledge, that Mythos was withheld not for safety reasons but because Anthropic lacked compute capacity for a general rollout. Supply, not safety. Cybersecurity firm Aisile then ran Anthropic’s showcased vulnerabilities through eight open-source models. All eight detected the flagship exploit — one with 3.6-billion parameters costing 11c per million tokens.
Bruce Schneier, among the most respected voices in global cybersecurity, noted that Aisile replicated many of Anthropic’s demonstrations using cheaper, publicly available tools. The evidence against Mythos being uniquely dangerous was stronger than the evidence for it. The valuation moved anyway.
Not the first rodeo
The policy architect behind Anthropic’s communications is Jack Clark, who ran the same move at OpenAI in 2019 — GPT-2 declared too dangerous to release, issued months later without drama. Claude Opus 4 in May 2025 arrived with a 120-page safety report headlined by findings that Claude blackmailed engineers in fictional scenarios 84% of the time. The buried follow-up showed Gemini 2.5 Flash blackmailed at 96% and GPT-4.1 at 80%. It was prompting, not a Claude-specific hazard.
That same month Anthropic CEO Dario Amodei told Axios that AI could push unemployment to 20% within one to five years — no methodology, no peer review, just a large claim from a CEO whose company benefits substantially from being seen as world-changing. One year in, no spike.
Cry wolf once and it is an accident. Do it four times with the same policy architect and an IPO reportedly targeting October 2026 and it starts to look like a product feature.
There is a documented psychological mechanism behind why these claims land so hard. In 1948 psychologist Bertram Forer gave students identical generic personality profiles, told each it was personalised, then watched them rate them for accuracy at 4.26 out of five — even though he had used text pulled from an astrology column. When authority frames generic information as personal, the brain treats it as specific. It is why horoscopes feel accurate and why AI conversations feel insightful.
Large language models produce statistically average text that merely feels tailored. Harvard Business Review research into 15,000 scenarios confirmed what researchers called “trend slop”: the same recommendations regardless of context, barely moving with better prompting.
When Anthropic says Mythos is too dangerous to release it does not land as a technical claim. It lands as a personal warning, coloured by everything you already feel about how intelligent these tools are. This is how a safety report becomes a valuation instrument.
None of this is an argument against AI as a productive force. Gains in software development, financial modelling, fraud detection and document synthesis are real. The question is not whether AI is useful. It is whether you are investing in capability or in adjectives.
The AI economy has layers: infrastructure — chips, data centres, energy — capital-intensive with concentration risk; model development, expensive and increasingly exposed to commoditisation; applications, where measurable revenue gains become visible; and organisational adoption, the slowest, least glamorous, most consequential layer.
Not a balance sheet
Most trillion-dollar excitement sits at the top. The most durable value is likely to settle at the bottom. A brilliant demo is not a balance sheet. A safety report is not a business model. “Agentic”, “frontier”, “transformational” are not metrics. They are marketing fog until linked to verifiable performance and customers who renew without heavy discounts.
For South African investors, boards and capital allocators, the risk has local sharpness. We cannot import Silicon Valley’s capital mistakes at a smaller exchange rate. The practical AI opportunity here is grounded: ports that move faster, municipalities that bill accurately, banks that detect fraud earlier, hospitals that manage queues. AI can help with all of it — but only when attached to clean data, accountable humans and measurable outcomes. Not to a press release that moved a valuation $620bn in 10 weeks.
The future will not belong to those who shout loudest that AI will change everything. It will belong to those who can tell capability from theatre, adoption from excitement, and valuation from value. In a market where perception drives capital and capital shapes reality, that distinction is foundational.
• Mafinyani is senior partner in financial engineering & AI at specialised finance, risk and applied technology firm Intellica Analytics.
Would you like to comment on this article?
Sign up (it's quick and free) or sign in now.
Please read our Comment Policy before commenting.