Anthropic’s New Update Challenges COBOL, Sparking IBM Stock Dive as US AI Giant Blames

Anthropic’s latest update reveals that its Claude model can generate “Claude Code” capable of automating COBOL modernization, a claim that sent IBM shares tumbling on the New York Stock Exchange, according to Bloomberg reporting via Zerohedge. The announcement, posted on Anthropic’s blog before 2 p.m. ET, highlighted that COBOL still processes roughly 95 % of U.S. ATM transactions and underpins “hundreds of billions of lines of code” in finance, airlines, and government systems. By offering AI‑driven code analysis and implementation, Claude promises to cut the cost and risk of updating legacy applications that have long suffered from a shrinking pool of skilled engineers. The market reaction was immediate: a wave of “panicked IBM longs” flooded the blog, and IBM’s stock fell sharply, underscoring investor anxiety over a potential disruption to the company’s legacy‑software services business.

The same update also coincided with Anthropic’s public accusation that three Chinese AI firms—DeepSeek, Moonshot AI and MiniMax—have been siphoning Claude’s capabilities through “distillation,” a technique that uses outputs from a more powerful model to boost a weaker one. The Guardian reports that Anthropic identified roughly 16 million interactions with Claude and 24 000 fabricated accounts used by the three labs to harvest functionality they had not independently developed. Anthropic argues that such illicit distillation bypasses U.S. export controls designed to preserve American dominance in advanced AI, and that models derived in this way lack the safety guardrails built into Claude to prevent misuse, such as restrictions on bioweapon design or cyber‑attack facilitation.

OpenAI, Anthropic’s chief rival, lodged similar complaints with U.S. lawmakers earlier this month, alleging that Chinese competitors were employing the same technique to erode the competitive edge of American chatbots. VentureBeat notes that Moonshot’s “Kimi K2 Thinking” and DeepSeek’s “V2.5” models have been touted as leading open‑source alternatives, but the reports do not confirm whether those models were trained using distilled Claude data. Nonetheless, the pattern of accusations suggests a broader geopolitical contest over AI intellectual property, with U.S. firms warning that unchecked copying could erode the strategic advantage conferred by their proprietary safety and alignment research.

Anthropic’s statement warned that the “window to act is narrow,” emphasizing the growing intensity and sophistication of the alleged campaigns. By publicizing the scale of the data theft—millions of queries and thousands of fake accounts—the company seeks to pressure regulators and industry partners into tightening enforcement of export rules. The Guardian adds that the practice poses national‑security risks, as models built from stolen outputs may be deployed without the rigorous testing that Anthropic applies to its own systems. If such models enter critical‑infrastructure environments, the lack of built‑in safeguards could expose vulnerabilities in sectors already dependent on legacy code, such as banking and aviation.

Investors appear to be weighing both the upside of Anthropic’s new COBOL‑automation capability and the downside of a potential escalation in AI‑theft disputes. IBM’s share price decline reflects fears that AI‑driven code generation could undercut the firm’s legacy‑maintenance contracts, while Anthropic’s move positions it as a front‑line defender of U.S. AI leadership. As the company rolls out Claude Code, market participants will watch whether the technology can deliver on its promise of cost‑effective modernization without triggering further regulatory scrutiny or retaliatory measures from Chinese AI labs.