Anthropic started utilising an early Claude Mythos Preview snapshot in February 2026 to identify security flaws in open-source software. This groundbreaking AI has been made with the specific purpose of conducting advanced cybersecurity operations such as identifying vulnerabilities in source code and the creation of exploit chains. This technology has increased its scope significantly since its inception. Anthropic announced in June 2026 that about 50 partners of the first Project Glasswing initiative were able to detect more than 10,000 security vulnerabilities with Mythos Preview, prompting Anthropic to add about 150 other organisations from over fifteen countries to the programme. Alongside the extended scope, the coordinated vulnerability disclosure dashboard at Anthropic reveals a major technological limitation, which is not related to discovering the vulnerabilities themselves, but rather to validating and disclosing them. In total, 2,300 vulnerabilities have been disclosed across 392 open-source projects, of which only 421 have already been fixed. This rapid development creates a critical legal dilemma: who owns knowledge of the discovered bug, and who can decide whether to keep the vulnerability confidential or disclose it?
INDIA IS ALREADY ENTERING THIS PROBLEM
India is also actively managing the duality of defensive and offensive AI. In terms of the former, the Reserve Bank Innovation Hub (RBIH) has created the MuleHunter, which is a system that uses AI and ML technology to detect mules as well as for fraud risk and anti-money laundering purposes. Meanwhile, the regulatory body is now formally recognising the offensive threat. In May 2026, CERT-In released its “Blueprint for Reducing Exposure and Defending against AI-Assisted Vulnerabilities Exploitation in Digital Infrastructure.” In June 2026, CERT-In issued its own guidelines on AI-Accelerated Vulnerability Protection and Response Requirements for Original Equipment Manufacturers and Technology Providers, indicating that India’s regulatory response is moving beyond basic preparedness to the specifics of managing AI-assisted vulnerabilities. By July 2026, the Government announced that CERT-In had carried out AI-assisted vulnerability assessments and cyber exercises targeting banking, finance, and insurance companies. India is dealing with both: AI which detects financial risks and AI that further discovers vulnerabilities in the financial sector.
WHAT EXACTLY IS THE IP IN AN AI-DISCOVERED VULNERABILITY?
Let’s take an example. The use of Mythos by a cybersecurity researcher results in the identification of an unknown vulnerability in the proprietary software of Bank X and leads to the creation of a new patch for that software. What will be the intellectual property in question?
A vulnerability cannot be per se an IP right. Rather, what can be protected as an IP may be knowledge of the vulnerability, exploit method, proprietary detection process, or even the technical solution created as a result. The copyright issues may arise as to the source code created by the AI, whereas patentability issues may arise as to the novel technical solution.
The protection of trade secrets becomes particularly crucial considering that the value of cybersecurity data could be based on secrecy. WIPO has identified three basic prerequisites for trade-secret protection, which include that such information should have commercial value as a result of its secrecy.
In India, no law in place provides comprehensive protection of trade secrets as an independent category. It mostly comes into play as an aspect of contractual duties, equity, and breach of confidentiality. In the case of American Express Bank Ltd. v. Priya Puri, the Delhi High Court has emphasised the significance of confidential information as opposed to information that is public knowledge or falls under the ambit of the employee’s general knowledge. This means that in India, whether the vulnerability discovered by AI can be protected or not would depend on its confidentiality.
This is further validated by recent studies in the literature. Fontanela, Costa and Marocco maintain that the use of trade secrets in protecting AI technology can promote both innovation and opaqueness, and hence digital vulnerability, when secrecy curtails transparency and auditing. The study focuses on AI technology in general and not the specific case of AI-discovered vulnerabilities.
THE REAL CONFLICT: SECRECY VS. VULNERABILITY
The biggest conflict emerges when the vulnerability that is detected by AI is something that impacts not only the entity itself but others as well. Consider the case where Bank A uses AI to detect a vulnerability which also exists at Bank B, the payment processors, and the government. Indeed, there is a good reason why things should remain secret. If they are revealed too soon, the vulnerability could be exploited before patches are released to fix the problem. In addition, revealing information will ruin the secrecy needed for trade secrets. Therefore, secrecy can be used in cybersecurity as well. The issue is that permanent secrecy produces the exact opposite situation. If Bank A is keeping the vulnerability secret while everyone else is still vulnerable to it, then trade secrets might end up being beneficial to Bank A but harmful to the security of the entire system. In CVD, the vulnerabilities are reported to the impacted organisation, verified and fixed before being disclosed publicly.
The vulnerability-disclosure program run by Anthropic is an example of this model in action. Discoveries made using Mythos are then triaged and reviewed prior to their being disclosed to the relevant maintainer, with details kept confidential until remediation is complete.Another unresolved problem regarding the adaptation of existing vulnerability-reporting practice in the case of AI is IP and ownership of vulnerability.
The key question is not merely whether the information regarding vulnerabilities should be kept secret. The real issue is for how long the information must be kept secret. The law of trade secrets protects secrecy as it generates economic value while cybersecurity laws necessitate disclosure as secrecy creates risk.
DOES AI CHANGE THE OWNERSHIP QUESTION?
The issue gets more complex when AI not only finds a weakness but also creates a new technical fix for it, raising the question of who the inventor is.
This reflects the same issue as in the DABUS case, where there was a dispute over whether AI could be recognised as an inventor. The issue becomes pertinent for India due to Section 3(k) of the Patents Act, 1970, which specifically excludes mathematical methods, business methods, computer programs per se, and algorithms from being patented.
Furthermore, even in cases where the invention developed is patentable, the prevailing patent system is built on the assumption that there will be a recognised human inventor. The dispute surrounding the use of AI as an inventor in India shows that the issue of AI inventorship is far from resolved. Although the 2025 CRI Guidelines published through the IP India website and the August 2026 Guidelines for the Use of AI in Patent Examination reveal growing engagement of institutions with AI, they fail to address the issue of inventorship of AI.
It shows how different these issues are. The bank can own information on vulnerabilities that is not patentable and not secret; similarly, for the bank to patent its technological solution would mean that it would have to reveal it.
INDIA’S POSSIBLE SOLUTION: CONTROLLED DISCLOSURE
India must therefore take into consideration a controlled-disclosure approach to AI-based vulnerabilities that impact critical infrastructure.
First, during the process of confidential discovery, the organisation must record the date of discovery, type of AI system used, human verification of the discovery, severity, and the impacted systems without any form of confidentiality. Second, during the process of protected disclosure, vulnerabilities must be communicated to the concerned party, including the CERT-In and relevant sector regulator, such as the RBI. Third, after remediation or expiration of reasonable disclosure period, public disclosure could follow.
It would not result in the elimination of the right to protect trade secrets, but in the acceptance of the fact that it is acceptable to keep information secret for the purpose of remediation, yet it is not supposed to become a method of suppression of the information that creates systematic risks forever.
CONCLUSION
India is utilizing AI systems like MuleHunter to curb financial crimes whereas emerging AI technologies such as Mythos have shown that they are capable of detecting weaknesses with machine speed. The issue then becomes not merely a matter of ownership of AI-generated cybersecurity information, but the time at which ownership gives way to responsible disclosure. Protection of trade secrets must continue for cybersecurity information which truly is confidential, but it cannot become an absolute shield in the case of an AI-generated vulnerability to interconnected financial infrastructure.
“With regards to the effect of frontier AI on cyber risk, which is seen by the Financial Stability Board (FSB) as being ‘the most immediate concern’ in August 2026,” this question is even more pressing. The right to keep an invention secret and at the same time disclose it can be a new paradigm for Indian IP law with the development of AI that can find its weaknesses itself.
Authored by: Ms. Shelly Jain
Shelly Jain is a B.A. LL.B. (Hons.) final year student at the Institute of Law, Nirma University, Ahmedabad. Her areas of interest include taxation, intellectual property law, and commercial laws.
