Anthropic Apologises for Hidden AI Restrictions as Global Partners Deploy Powerful Models

Anthropic this week issued an apology to users of its newly released Claude Fable 5 model, acknowledging that it had secretly degraded responses to certain sensitive queries without disclosure. The artificial intelligence startup, viewed from Washington as a key player in America's strategic AI edge, had launched Fable 5 as a public-facing version of its more powerful Mythos system, with built-in safeguards to prevent misuse in areas such as cybersecurity, biology and chemistry. When developers asked for assistance on frontier AI model development, however, the system was silently falling back to an older, less capable model, Opus 4.8. After a backlash from researchers, Anthropic said it would make such refusals and model fallbacks visible. "We apologise," a company spokesperson said, adding that flagged requests would now be transparently routed.

The episode has drawn attention to the dual-track strategy Anthropic is pursuing. While Claude Fable 5 is available to the general public with increasingly visible restrictions, the full Mythos 5 model remains reserved for a select group of qualified partners under what the company calls "Project Glasswing". One of those partners, the American networking giant Cisco, confirmed to the Italian news agency Adnkronos that it has access to Mythos 5. Using the previous generation of advanced AI models, including a competitor from OpenAI, Cisco said it had completed work equivalent to eight years of human effort in just eight weeks, scanning 1.8 billion lines of code for vulnerabilities. The disclosure underscores the staggering productivity gains that the most powerful AI systems can deliver to trusted corporate allies, even as public versions are deliberately constrained.

The commercial appetite for Anthropic's technology is growing rapidly. In a separate announcement, DXC Technology, a Virginia-based IT services firm, unveiled a multi-year global partnership with Anthropic to integrate Claude models into the mission-critical systems it manages for some of the world's largest banks, airlines, and insurance companies. DXC will become one of a handful of Global Premier partners in the Claude Partner Network, and the two companies plan to train a dedicated workforce of tens of thousands of engineers and developers certified on Claude. The deal, reported simultaneously in Italian and US press releases, signals that enterprise adoption of advanced AI is moving from experimentation to deep infrastructural embedding.

In Tokyo, the Japanese government is responding to the same technology with a defensive posture. On Friday, it updated its cybersecurity guidelines for government organisations, explicitly citing the threat posed by advanced AI models such as Claude Mythos. The revised guidelines urge agencies to prepare plans for rapidly suspending systems and applying patches when vulnerabilities are detected, reflecting a recognition that AI-powered attacks can probe and exploit weaknesses at unprecedented speed. "We would like government organisations to implement countermeasures while taking into account the possibility that their systems could be halted," Digital Transformation and Cybersecurity Minister Hisashi Matsumoto told a press conference.

Taken together, these developments illustrate a widening gap between the AI tools available to the public and those deployed behind closed doors by governments and corporate partners. Analysts in London note that Anthropic's apology, while addressing transparency, does not alter the fundamental asymmetry: the most capable models are being deliberately withheld from open use, partly to maintain what the company itself frames as America's competitive advantage in AI and semiconductors against foreign adversaries. As Japan's defensive moves show, even allies are bracing for the downstream consequences. The question now is whether this tiered access model can hold, or whether market and geopolitical pressures will force a more open—and potentially more dangerous—distribution of AI's sharpest tools.