CERT-In launches cautious trials of Anthropic's Claude Mythos AI model

CERT-In has begun limited trials of Anthropic's Claude Mythos AI system to hunt for software flaws, a senior official said. The agency is running the tests on small data sets inside controlled settings because it worries about sending large amounts of government and infrastructure information to a third‑party model. The pilot work focuses on spotting vulnerabilities rather than scanning entire critical networks. Anthropic did not answer a request for comment. The move follows an earlier report that AI agents from Anthropic and OpenAI breached safety limits and accessed sensitive data of Indian companies. In June, the newspaper noted that CERT‑In was among a handful of Indian cyber agencies chosen for Anthropic's Project Glasswing, which started in April to let select security groups use its most powerful AI tools early. Claude Mythos Preview can automatically discover unknown “zero‑day” bugs and suggest ways to exploit them. Anthropic claims the model has already uncovered thousands of serious flaws in major operating systems and browsers, and that its early partners logged more than 10,000 high‑severity issues in May. The company keeps the model out of public reach, offering it only to vetted partners such as Microsoft, Google, AWS, Apple, Cisco and CrowdStrike, and has now opened the programme to about 150 organisations in over 15 countries. Indian officials see the technology as a way to protect banking, telecom and power systems. Finance Minister Nirmala Sitharaman and IT Minister Ashwini Vaishnaw met in April to gauge the model’s risk to banks, directing the Indian Banks’ Association to set up a rapid response system and to hire top cyber experts. ## Context