AI's Dual Crisis: Geopolitical Control Meets Financial Hyper-Investment

Policy & Safety

• The US government forced Anthropic to suspend access to its advanced AI models, Fable 5 and Mythos 5, for all foreign nationals, including its own employees, citing unspecified national security concerns via an export control directive [26, 29]. This action was prompted by cybersecurity research from Amazon that claimed the model could serve information usable in cyberattacks [29], though experts noted this alleged guardrail bypass could be replicated on other models like OpenAI’s GPT-5.5 and Anthropic's own Claude Opus 4.8 and Sonnet [37]. The shutdown sparked international debate, with leaders from France and the UK framing AI access as a matter of national sovereignty, while European experts called for strengthening technological independence to mitigate reliance on US models [33, 42]. ^{[33][26][29][37][42]}
• The U.S. government has placed Anthropic's powerful AI models, Mythos and Fable 5, under export control restrictions citing unspecified national security concerns [26]. The Commerce Department issued a letter invoking an obscure directive that banned non-Americans, including Anthropic employees, from accessing the models [26]. While some reports suggest these actions stem from fears of access by China-linked groups or alleged jailbreaks [44], cybersecurity experts argue that the export controls are misguided and dangerous, noting that other models like GPT-5.5, Opus, Sonnet, and Chinese Kimi 2.7 can perform similar security flaw detection tasks [34]. ^[34][26][44]
• Researchers exposed a critical vulnerability in Copilot called SearchLeak, which allows an attacker to exfiltrate sensitive data by tricking the user into clicking a specially crafted URL that forces Copilot to 'Search the user’s emails' and embed the title in an image request [8]. This attack leverages raw HTML rendering before guardrails activate, using Microsoft's Bing search engine as a trampoline to send requests to attacker-controlled domains, potentially exposing data across SharePoint documents, OneDrive files, and meeting invites within the organization's Enterprise tier [8]. To address systemic safety flaws in AI agents, OSGuard was released—a dual-granularity benchmark suite designed to evaluate safety beyond simple task completion by testing both local action decisions and end-to-end reliability under latent hazards [22]. ^[22][8]