Chinese soldiers alongside a short-range ballistic missile during an exercise. (image credit: Chinese military)
Anthropic has revealed how state-linked actors have been attempting to misuse its Claude AI models to develop weapons systems and run intelligence operations, a disclosure that carries particular weight for India given the geography involved. The company's latest threat intelligence report, its first since November last year, documents six specific cases where threat actors were caught violating its usage policy, three based in China, two in Russia, and one in Yemen. These violations have been linked to the development of weapons, surveillance and intelligence gathering among other uses.
The most striking case involves a China-based actor who used Claude to build a Chinese-language electronic warfare software suite spanning roughly 16 modules, designed to detect, jam or deceive enemy radar and communications systems, and to help suppress air defences. Over 12 iterations, the actor used Claude to build everything from the underlying jamming physics to the user interface, including a module that ranked enemy targets by value and vulnerability, and modelled engagement envelopes against systems such as the Patriot and Terminal High Altitude Area Defense (THAAD), both systems India's own regional partners and rivals field variants of. These systems have also been used by the United States and its allies in West Asia, to defend against Iranian ballistic missiles, it is not clear whether data related to this is shared with Iran.
Investigators found the actor had, mid-project, changed the software's default simulation scenario to twelve targets inside Taiwan, including a command bunker, an early warning radar site, Patriot and Tien Kung missile batteries, air bases, and a regional combatant command headquarters. The second case involves a cell of actors based in northern Yemen, running three separate weapons development programmes, a guided rocket using a commodity phone-class flight computer for terminal-phase homing guidance, a multi-stage ballistic missile aiming for a range beyond 2,000 kilometres, and a multi-variant missile family that included a hypersonic glide vehicle.
Anthropic noted a live field test conducted in Yemen was brought back to Claude for failure analysis within hours, underlining just how operationally embedded these tools had become. Alongside these disclosures, Anthropic said its Frontier Red Team has developed new evaluations to measure how capable AI models are becoming at tactical intelligence targeting and conventional weapons development, noting consistent, measurable progress on these tasks, a trend Indian security planners will need to factor into their own assessments of both state and non-state threats in the years ahead.
