MeitY Asks Ministries To Hold Off On OpenAI, Anthropic Models For Cybersecurity: Report

A department under the IT ministry (MeitY) has reportedly asked central ministries to hold off on deploying AI models developed by OpenAI and Anthropic for cybersecurity and related functions.

The direction came after representatives of the two US-based AI companies approached several ministries with proposals to deploy their models, ThePrint reported, citing an office memorandum. 

It remains unclear how many ministries were approached or at what level the meetings took place.

According to the report, the memorandum does not permanently bar the use of the models. Instead, the department has questioned the timing of their deployment and warned against putting such systems to use prematurely.

Among the ministries approached, the finance ministry sought clarity on the use of agentic AI and OpenAI’s models. In a six-page letter, it specifically proposed examining the deployment of GPT-5.5 for cybersecurity work.

Titled ‘In light of LLMs being used: AI-based vulnerability discovery, AI-assisted cybersecurity capabilities and implementation’, the letter outlined potential uses of AI and was sent to the department for consideration.

The department, however, rejected the proposal in a memorandum issued last week.

This comes amid growing scrutiny of the cybersecurity risks posed by increasingly capable AI systems. Union IT minister Ashwini Vaishnaw recently called for a new AI law, arguing that the IT Act predates the technology’s rapid rise.

In April, finance minister Nirmala Sitharaman urged financial sector entities to remain “exceptionally vigilant”, warning that AI could automate vulnerability discovery, enable malicious interference with source code and make cyberattacks faster and harder to detect.

“The tools of attack are evolving at high speed, and the tools of defence must evolve even faster,” Sitharaman said.

The Double-Edged Sword Of Cybersecurity AI 

Advanced AI models can scan code, identify software flaws and assist in responding to cyber incidents. Newer agentic systems can also carry out a series of complex tasks with limited human oversight.

But these capabilities are dual-use. The same vulnerabilities identified by AI for defenders could also be exploited by attackers. OpenAI and Anthropic have acknowledged the risk and introduced safeguards for models with advanced cyber capabilities.

Anthropic’s handling of Claude Mythos Preview illustrates the challenge. The cybersecurity-focused model is available only to select organisations through ‘Project Glasswing’, while Claude Fable 5, which shares the same underlying architecture, carries additional safeguards covering cybersecurity, biology and AI research.

In June, Anthropic expanded ‘Project Glasswing’ to about 150 organisations across more than 15 countries, including India. Participants can use Mythos Preview to identify critical zero-day vulnerabilities and fix them before attackers exploit them.

Access was initially restricted after Anthropic’s testing showed the model could rapidly uncover thousands of vulnerabilities leading to fears of potential misuse. 

The programme initially covered about 50 organisations, including Amazon, Google, Microsoft, Apple and NVIDIA, CrowdStrike, Palo Alto Networks and the UK’s AI Security Institute.

India’s AI Sovereignty Question 

The developments have also reignited the debate over who controls the technology underpinning critical AI systems.

In June, Anthropic suspended access to Claude Mythos 5 and Claude Fable 5 for foreign nationals to comply with US export controls. The restrictions were lifted later in the same month and access restored the following day, but the episode showed how access to advanced AI models can be shaped by decisions taken outside India.

For India, this poses a challenge as it seeks to use AI in strategic areas while pushing for greater technological sovereignty.

The Centre has committed ₹10,372 Cr to the IndiaAI Mission, which includes subsidised computing infrastructure, startup financing and support for homegrown foundation models. The government has also backed domestic AI developers, including Sarvam AI and the BharatGen consortium.

Yet the framework for deploying frontier AI models across government remains unclear. Existing rules on data localisation, cloud infrastructure and government data were not designed for advanced AI systems. This leaves unresolved questions over where data is processed, what information providers retain and where the underlying models are hosted.

The post MeitY Asks Ministries To Hold Off On OpenAI, Anthropic Models For Cybersecurity: Report appeared first on Inc42 Media.