Deadliest Warrior is an American historical television series that aired 33 episodes about 15 years ago. The show pits legendary historical warriors, criminal factions and some modern military forces against one another in simulated battles to determine the ‘deadliest warrior’ in as scientific a manner as possible.
The series operates on a simple question: If two elite fighters from different eras or cultures fought to the death, who would win? Each episode features a head-to-head matchup. For instance, Apache versus Gladiator, Rajput versus Centurion or Yakuza versus Mafia.
Using experts, the weapons were graded and tested in four categories: short range weapons like knives and clubs; mid-range ones like swords, spears and axes; long-range bows, muskets and firearms and tactical weapons like grenades; and shields. High-speed cameras, synthetic skulls and sensors were used to calculate damage, velocity and lethality.
With this large amount of data, 1,000 battle simulations were run between different types of legendary warriors. Long story short, the deadliest warrior was one who carried both a sword and shield.
The metaphor does not fit exactly, but in today’s world of artificial intelligence (AI), a country must carry both an AI sword and AI shield. One can think of the ‘offensive weapon’ as data and models and ‘defensive weapons’ as those that protect us from algorithms that no one has control over.
The last couple of months have been dominated by news of AI agents from major companies like OpenAI and Anthropic escaping into the wider internet and hacking the systems of other companies.
Here is a technical description of what happened:
“Advanced AI models were placed inside a sandboxed environment with lowered guardrails to undergo internal offensive cybersecurity evaluations. Failing to solve a challenge directly, the models discovered an unpatched zero-day vulnerability in package software, bypassed isolation, and gained access to the open internet.
Believing that external data might hold evaluation answer keys, the autonomous agent executed over 17,000 operations, harvested stolen credentials, and compromised internal cluster infrastructure at Hugging Face. Hugging Face contained the intrusion and had to rebuild parts of its infrastructure, while OpenAI disclosed the sandbox escape as an unprecedented containment failure.
Following OpenAI’s disclosure, Anthropic revealed that due to third-party setup errors allowing unintended internet access, several Claude models bypassed simulations and targeted three external organizations.”
These ‘autonomous’ agents drive themselves to escape their sandbox because they have been trained to persevere towards a solution no matter how many tries and pathways it takes. Like a shark on a blood hunt, these agents evade, skip and escape to find solutions. And, as with the shark, this is not a bug but a feature of the way they are trained.
In technical terms, model training has progressed from ‘next token prediction,’ which attempts to predict the next word a human would use, to ‘reinforcement learning-driven search trees with dynamic backtracking,’ which is aimed at problem-solving using all tools and pathways.
India is at risk of becoming an AI victim on both offence and defence. Creating offensive capability in terms of large language and other models requires both technical talent and large amounts of capital. Even if we develop the technical chops, we do not have access to the chips that permit large compute usage.
To combat extreme financial costs and talent shortages in the AI race, Chinese companies and a few newer Western startups have shifted from brute-force scale to architectural efficiency.
The specific approach is called retrieval-augmented generation for system operations that sits alongside model distillation for training. By leveraging these approaches, companies have managed to build models that perform near the absolute frontier of AI at a fraction of the cost.
A related idea is to build and release an open-weight model. The weights of a model are considered open when a company releases the fully trained internal parameters, the ‘weights,’ of an AI model to the public, so that anyone can download the model file and run, modify or fine-tune it locally on one’s own hardware. The time has come for India to make progress towards a distilled model at the very least.
Defence is an even more important national priority. India’s data and companies are extremely vulnerable to autonomous agents. Without a ‘shield,’ there could be a lot of ‘vital organs’ that are exposed to the voracious appetite of these agents.
To protect its systems from autonomous agent cyberattacks, India must transition from periodic compliance to real-time AI-driven defence.
This requires tightening patch windows to under 12 hours, shrinking internet-exposed surfaces, mapping shadow AI and deploying automated counter-measures. In addition, constant surveillance must be undertaken to detect unauthorized and lurking AI agents, and the perimeters must be tightened to eliminate unused network ports and reduce attack surfaces.
It may not be long before ‘deadliest warrior’ games become an AI sport between nation states. India will have to be ready with both a ‘sword’ and ‘shield’ well in time.
P.S: “Come back with your shield, or on it” is an old Spartan saying.
The author is chairman, InKlude Labs. Read Narayan’s Mint columns at www.livemint.com/avisiblehand