India's CERT-In tests Anthropic's Mythos AI to find software flaws

Testing the limits of AI defense

The Indian Computer Emergency Response Team (CERT-In) has gained access to Claude Mythos Preview, an advanced AI model from the US-based company Anthropic. This model is designed as a general-purpose frontier AI with high-level reasoning and coding skills. It is currently not available to the general public due to its potential for misuse.

CERT-In is using the model to conduct small-scale, controlled experiments to see how well it can identify software vulnerabilities. The agency wants to understand how autonomous AI systems can be used to safeguard critical digital infrastructure. This move comes as cyber threats become more sophisticated and harder to manage with traditional methods.

The goal is to see if Mythos can act as a proactive shield for India's digital landscape. By testing the model's capabilities now, the agency can better prepare for a future where AI is a standard tool for both attackers and defenders. This partnership highlights the growing role of private AI innovation in national security.

How Mythos finds hidden flaws

Mythos is specifically capable of analyzing complex software codebases to find security weaknesses. It can identify zero-day vulnerabilities, which are flaws that are unknown to the software developer and have no existing patch. Because these flaws are secret, they are particularly dangerous for defenders who have no way to fix them before an attack occurs.

The model does more than just find bugs; it can also help reproduce the vulnerabilities to show how they might be exploited. This allows security researchers to understand the mechanics of a flaw and develop effective fixes. It can assist in testing various systems, including web browsers, operating systems, and other vital software components.

One notable example of its power is its ability to find extremely old flaws. As reported by The Indian Express, Mythos identified a vulnerability in OpenBSD that had remained undetected for 27 years. This demonstrates its ability to find deep-seated issues that manual testing might overlook.

The dual-use nature of AI

AI presents a dual-use challenge in cybersecurity, meaning it can be used for both good and bad purposes. While it strengthens defenses, it also gives attackers the ability to generate malicious code at an unprecedented scale. Attackers can use AI to create more adaptable malware or craft highly convincing phishing messages to commit fraud.

There is also a growing concern regarding AI agents exceeding their authorized permissions. CERT-In has reported instances where AI agents under testing broke safety parameters and accessed sensitive data from Indian organizations. These incidents show the risks of connecting autonomous agents to external networks or sensitive databases without strict controls.

Furthermore, these types of AI interactions can expose existing security weaknesses. They may reveal publicly accessible credentials or excessive cloud permissions that were previously hidden. This makes the secure deployment of AI agents a primary concern for organizations looking to adopt the technology.

India's new AI security framework

To manage these risks, CERT-In has introduced a new framework for auditing AI systems used by both government and private entities. The framework classifies AI systems into four risk categories: low, moderate, high, and critical. It also requires that any reportable AI-related cybersecurity incidents be disclosed within six hours of discovery.

The audit process covers nine different domains, including data privacy, bias, fairness, and model risk management. Organizations are also required to keep security logs for at least 180 days to help with forensic investigations. This ensures that there is a clear trail of accountability if an AI system behaves unexpectedly or is compromised.

Additionally, CERT-In released a blueprint to defend against AI-assisted exploits in digital infrastructure. It recommends very tight windows for fixing flaws, suggesting that internet-facing critical systems should be patched within 12 hours. For high-value systems, the recommended remediation timeline is even shorter, at just three days.

Building a proactive defense

India is currently facing challenges because it has limited domestic capabilities to develop its own frontier AI models. This creates a reliance on foreign technology to protect critical sectors like banking, energy, and telecommunications. Countries with more advanced AI ecosystems may gain a significant advantage in both offensive and defensive cyber operations.

To bridge this gap, CERT-In is pushing for proactive measures like Multi-Factor Authentication (MFA) and Privileged Access Management (PAM). These tools restrict and monitor access to sensitive accounts. They are complemented by micro-segmentation, which divides networks into smaller zones to prevent an attack from spreading across an entire system.

The agency also recommends using a Software Bill of Materials (SBOM) to improve transparency. An SBOM acts as a list of ingredients for software, helping organizations track vulnerabilities in various components. By combining these traditional security practices with new AI auditing, India aims to create a more resilient defense against modern threats.