Detailed Analysis
India's Computer Emergency Response Team (CERT-In) has issued a high-risk advisory warning Indian firms and micro, small, and medium enterprises (MSMEs) about emerging cyber threats posed by advanced AI systems, specifically citing Anthropic's Claude Mythos as a paradigm-shifting risk. The advisory urges organizations to implement proactive defensive measures in response to AI-powered attack capabilities that require minimal human effort to execute. CERT-In's guidelines emphasize rapid remediation of known vulnerabilities, improved asset inventories, and resolution of fragmented ownership structures within organizations — structural weaknesses that become acutely dangerous when adversaries can leverage tools capable of autonomously discovering and exploiting security flaws at scale.
Claude Mythos, also referred to as Claude Mythos Preview, is an experimental Anthropic model developed explicitly for cybersecurity research rather than general public use. Its capabilities are substantial and well-documented: the system can autonomously scan both open- and closed-source software for zero-day vulnerabilities, generate working exploits from discovered flaws, reverse-engineer proprietary software, and simulate multi-step attack chains — including a demonstrated 32-step enterprise attack sequence in controlled testing. Independent and Anthropic-conducted assessments found Mythos capable of identifying a 27-year-old vulnerability in OpenBSD, surfacing thousands of high- and critical-severity issues across major operating systems and browsers, and contributing to over 270 patches in Firefox alone. Its success rate on expert-level cyber tasks has been measured at approximately 73%, representing a meaningful leap over predecessor models like Claude Opus.
The urgency of CERT-In's advisory is compounded by a concurrent security incident at Anthropic itself. The company is actively investigating unauthorized access to Claude Mythos through a third-party vendor, a breach attributed to what appears to be a combination of insider threat exposure and operational security failures rather than a compromise of Anthropic's core infrastructure. Critically, no confirmed misuse — such as the execution of offensive cyber prompts — has been documented at this time. Anthropic had previously restricted Mythos access to select cybersecurity firms precisely because of concerns that broader availability would empower malicious actors, making the third-party access incident a pointed validation of those fears and a cautionary signal for any organization in the AI supply chain.
The broader context of CERT-In's warning reflects an accelerating global dynamic in AI-enabled offensive security. The dual-use nature of tools like Mythos — simultaneously valuable for defensive research and potentially devastating in adversarial hands — is forcing regulatory and governmental bodies worldwide to grapple with frameworks that existing cybersecurity law was not designed to address. For Indian firms and MSMEs in particular, many of which operate with limited security staffing and slow patch cycles, the advisory underscores a structural vulnerability: even well-resourced organizations that are slow to remediate known weaknesses face compounded risk when an adversary can automate vulnerability discovery and exploit generation in hours rather than weeks. CERT-In's intervention signals that AI-specific cyber threat advisories are becoming a standard instrument of national cyber defense posture.
Anthropic's approach to Mythos — controlled access, delayed public release, and active investment in its own safety assessments — represents one model for how frontier AI labs might navigate the tension between advancing offensive security research and preventing harm. However, the unauthorized access incident illustrates that even well-intentioned access controls are only as strong as the weakest link in a vendor ecosystem. As AI systems grow more capable of compressing the time and expertise required to conduct sophisticated cyberattacks, the incident and CERT-In's response together point toward an emerging norm: AI cybersecurity models will increasingly be treated not merely as software products but as critical infrastructure requiring the same governance rigor applied to weapons-adjacent technologies.
Read original article →