CERT-In and Claude Mythos: How AI Can Find Software Flaws

Controlled trial: The Indian Express cited a senior official saying that CERT-In received access to Claude Mythos and was testing it on smaller datasets in controlled environments. The account points to limited vulnerability research. The report did not establish wider operational deployment or identify the precise model version. Mint had earlier reported on CERT-In’s evaluation of other AI models for finding weaknesses, putting the Mythos trials in the context of a broader testing effort.

Project Glasswing: Anthropic launched the initiative to give vetted organisations access to advanced models for defensive security work. Its launch announcement concerned Mythos Preview. Anthropic subsequently described expanded access and other Mythos releases, making it important to distinguish the programme from any specific model version.

The weakness: A software vulnerability is a flaw that can compromise a system’s confidentiality, integrity or availability. A zero-day vulnerability is commonly understood as a flaw unknown to the responsible vendor or defenders when discovered or exploited, leaving them with no prior time to prepare a fix.

Exploit and patch: An exploit is a method or piece of code that takes advantage of the flaw. A patch is a software change that fixes it. A report from a model still needs human verification. Finding a weakness, showing that it can be exploited and releasing a repair are three distinct steps.

Incident response: Under Section 70B of the Information Technology Act, 2000, CERT-In is the national agency for cyber incident response. Its functions include gathering incident information, issuing alerts and advisories, and coordinating response. Its 2022 directions require covered organisations to report specified incidents within six hours of noticing them or being informed of them.

Critical systems: The National Critical Information Infrastructure Protection Centre, or NCIIPC, is the national nodal agency under Section 70A for protecting critical information infrastructure. The term covers computer resources whose disruption could have a debilitating impact on national security, the economy, public health or public safety.

Defensive gains: AI can review large codebases, prioritise suspicious patterns and suggest tests. A defender can confirm the result in a controlled environment and share it privately with the maintainer for repair.

Misuse potential: Similar capabilities can help attackers locate weak systems. A tool-using model may also follow malicious instructions embedded in material it reads or make an unintended network request if permissions are too broad. Model-generated findings can contain errors and should be independently reproduced.

Sensitive inputs: Production logs and source code may contain credentials, personal data or details of critical systems. External processing raises questions about what data leave the organisation, who can access them, where they are processed and for how long they are retained.

Containment: Use synthetic or sanitised samples, restrict network access, provide only the permissions needed for the task, log model actions and require human approval before interacting with real systems. For confirmed vulnerabilities, coordinated disclosure gives maintainers time to prepare fixes before technical details spread.

Validation and disclosure: An analyst first reproduces the finding against the affected software version in an authorised test environment. If it is real, the analyst records which versions are affected and what conditions an attacker would need. Under CERT-In’s responsible disclosure policy, CERT-In validates reports and can coordinate privately with the researcher and vendor. The vendor can then develop a fix or workaround, and CERT-In can issue an advisory at an appropriate time.

CVE identifier: Common Vulnerabilities and Exposures gives publicly disclosed vulnerabilities standard identifiers. A CVE ID lets different teams refer to the same flaw across advisories and asset records. A patch is only useful to an organisation once it has identified the affected installations and applied the fix or deployed an effective workaround.

Measure the findings: Test the AI and a conventional scanning workflow against the same set of authorised systems. Compare the number of independently confirmed discoveries, false alerts and analyst hours needed for each finding. Record whether fixes are applied faster. Counting model alerts alone gives little indication of how much risk was reduced.

Prioritise by exposure: A flaw in an internet-facing service or a system that keeps a hospital or power network running may require urgent containment. Teams should assess whether attackers can realistically exploit the flaw, what service could be affected and whether a patch is available. This connects an AI experiment with the operational responsibilities of CERT-In, system owners and sectoral teams.