New Delhi: A department under the Ministry of Electronics and Information Technology (MeitY) has asked ministries not to deploy OpenAI and Anthropic’s models for cybersecurity and related functions for now, ThePrint has learnt, days after the two companies approached several ministries with their proposals.
An office memorandum has been circulated to this effect, asking ministries not to put such models into use prematurely. The representatives of the two companies met officials across ministries to press for deployment, a source told ThePrint.
The number of ministries approached and the level at which the meetings took place could not be confirmed.
The Ministry of Finance was among those that wrote to the department, asking if the ministry could examine the use of agentic AI and the deployment of OpenAI models. In its six-page letter, reviewed by ThePrint, the Ministry of Finance specifically asked about deploying GPT-5.5 for cybersecurity work.
The letter, titled ‘In light of LLMs being used: AI-based vulnerability discovery, AI-assisted cybersecurity capabilities and implementation’, set out proposals on what could be deployed and was sent for consideration.
In a memorandum last week, the department returned a firm no.
The source quoted above said the memorandum does not amount to a permanent restriction. The department’s objection is about the timing of deployment rather than a bar on the technology.
The models in question
Both companies market their frontier models for defensive cybersecurity: automated vulnerability discovery, in which a model scans codebases for exploitable flaws; code review; incident triage; and agentic deployment, where a model executes multi-step tasks with limited human supervision.
The capabilities are dual-use. A model that finds a vulnerability in defensive testing can find one for offensive purposes, and both companies acknowledge as much in their published safety frameworks, which assign cyber capability thresholds to models and set restrictions accordingly.
Anthropic’s most capable models sit in a tier it calls Mythos. Claude Mythos 5 and Claude Fable 5, released 9 June, share an underlying model; Fable carries additional safeguards for biology, cybersecurity and AI research. Fable is publicly available. Mythos Preview is not, and is restricted to a small number of organisations under a programme the company calls Project Glasswing.
On 12 June, Anthropic suspended access to both models to comply with export controls imposed by the US Department of Commerce. The controls were lifted on 30 June, and access was restored the following day, according to a statement published by the company.
The sovereign AI context
The advisory comes as the government pushes a sovereign AI agenda. The IndiaAI Mission, approved by the Cabinet in March 2024 with an outlay of about Rs 10,372 crore, funds compute capacity, datasets and indigenous foundation models. Under its foundation model track, MeitY has backed domestic developers including Sarvam AI and the BharatGen consortium.
The objective is to reduce dependence on foreign models for functions the government treats as strategic. Cybersecurity is one. CERT-In operates under MeitY, and critical infrastructure protection is coordinated by the National Critical Information Infrastructure Protection Centre under the National Technical Research Organisation.
India is among the largest user bases for both OpenAI and Anthropic. Both companies have expanded in India over the past year, OpenAI with an office in New Delhi and Anthropic with one in Bengaluru.
Government procurement of AI systems remains without a settled framework. Rules on data localisation, cloud empanelment and the classification of government data constrain what can run on foreign infrastructure, but they were not written for model deployment. They do not address where inference takes place, what logs a vendor retains, or whether model weights are held outside India.
(Edited by Ajeet Tiwari)
Also Read: Why Anthropic and Open AI want a global AI watchdog