#105 AI Agents Take the Stage
Today, Bharath Reddy talks about new developments pertaining to AI models, followed by Avinash Shet who sheds light on India’s innovation capabilities. Lastly, in our newly introduced section that curates reading recommendations, this week Lokendra Sharma looks at the issue of arms control in cyberspace.
Technopolitik: The Next AI Frontier?
— Bharath Reddy
The language generation abilities of AI models that wowed us when OpenAI launched chatGPT not too long ago are now taken for granted, and we also have a much better sense of their abilities and limitations. The ability to train on massive amounts of data and generate intelligent-sounding text is unlikely to be the path to general intelligence. However, AI agents like the one Anthropic has just announced API access for could open up a whole new world of possibilities (and, of course, risks).
Anthropic, the company behind the frontier AI model Claude, has launched public beta access to computer use. Developers can direct Claude to use computers the way people do through API access. Here’s how the company describes it:
“Available today on the API, developers can direct Claude to use computers the way people do—by looking at a screen, moving a cursor, clicking buttons, and typing text. Claude 3.5 Sonnet is the first frontier AI model to offer computer use in a public beta. At this stage, it is still experimental—at times cumbersome and error-prone. We're releasing computer use early for feedback from developers, and expect the capability to improve rapidly over time.”
There has yet to be a consensus on what an agentic AI system is. In a recently published report, researchers at the Centre for Security and Emerging Technology (CSET) describe four characteristics of increasingly agentic AI systems: “They pursue more complex goals in more complex environments, exhibiting independent planning and adaptation to directly take actions in virtual or real-world environments”.
The CSET report presents a few governance implications:
Current assessments of their capabilities and real-world impacts are limited, highlighting the need for methodologies to track their progress
Technical guardrails to enable thoughtful design of agents that prioritise security, privacy, visibility, trustworthiness, and control.
Many existing laws provide legal guardrails in how AI agents are governed. However, challenges will still remain regarding AI consciousness, legal personhood, negligence standards, and principal-agent relationships.
What does it mean for India? Nandan Nilekani has stated that India’s focus need not be on building another LLM but on being the use case capital of the world. Takshashila’s paper on AI governance also makes similar observations. Access to top-tier research talent and massive investments in computing would be a barrier to building frontier AI models. However, with vast experience in software development and access to engineering talent, Indian companies can have an advantage in building applications on top of such models. For instance, Anthropic claims that Asana, Canva, Cognition, DoorDash, Replit, and The Browser Company have already begun to explore the possibilities of AI agents and we can imagine many Indian companies addressing various such use cases.
There are bound to be many instances where things go wrong when AI agents are deployed. The capabilities of AI agents have been compared to those of self-driving cars at the moment, where consumers can’t really trust it to drive on its own and need to keep their hands on the wheel. Developing methodologies for understanding and assessing the progress of AI agents and building the right governance guardrails are the priorities to ensure they are deployed responsibly.
Technomachy: Is India the Next Innovation Hotspot? A GII Analysis
— Avinash Shet
The world is changing at a rapid pace in recent decades. Innovation and technological advancements have resulted in a boom in economies, the elevation of people from poverty, and increased life expectancy and lifespan. Technological superiority is playing an important role in shaping the world’s political dynamics and geopolitics.
India is at the forefront of building its innovation ecosystem by leveraging technology. This can be witnessed with India jumping up the rank year after year at the Global Innovation Index (GII), which is published by the World Intellectual Property Organisation (WIPO). The GII assesses the innovation capabilities of nations worldwide.
This year, India secured 39th rank, jumping 37 ranks from last decade. It is the leading nation in the Central and Southern Asia region and the lower-middle income group. The GII analyses a couple of factors: innovation input and output. Innovation input evaluates a country's institutions, human capital, research, infrastructure, market sophistication, and business sophistication, while innovation output assesses knowledge and technology outputs, as well as creative outputs. According to the index, India ranks higher in innovation output (33rd rank) than in input (44th rank). India has demonstrated strengths in knowledge and technology outputs, as evidenced by its patents, scientific publications, high-tech manufacturing, intellectual property, and high-tech exports, among other areas.
Understandably, India is ranked one in Information and Communications Technology (ICT). Surprisingly, unicorn valuation is ranked 9th under innovation output, which showcases India’s budding startup landscape. Other metrics, such as startup finance and scaleup, which rank 8th, and venture capital value, which rank 6th, support the claim.
This is where the positive part ends. The remaining data show us the gaps India has to work toward. India ranks 72nd in infrastructure; surprisingly, under this category it is ranked 110th in ICT access, 42nd in the government's online service, and 61st in e-participation. This is despite India's efforts to invest in digital public infrastructure (DPI) and its innovation. India also didn’t do well in ecological sustainability (ranked 97th), even though the government strives to achieve its sustainability goals.
India has consistently grappled with the task of transforming its education system. Despite the policies in place like the National Education Policy (NEP), India still faces significant challenges. The GII ranking also reflects this, with India ranking 82nd in education. The GII ranking reveals an unexpected trend in the research and development (R&D) ecosystem. Despite investing only 0.6% of its GDP in R&D, India outperforms 79 other countries, ranking 54th. The dealbreaker information that goes against our expectations is global corporate R&D investors. With only 40% of total R&D investment in India coming from the private sector, the 18th rank is quite surprising.
According to the QS University ranking 2025, the Indian Institute of Bombay is ranked first in India and 118th globally. Even with this stat, India ranks 24th in the QS university ranking subcategory in GII.
The GII ranking provides intriguing insights on India’s status as an aspirant innovation hub. The GII ranking places India, a lower middle-income country, among the higher and upper-middle-income countries. It has performed a lot better in the startup ecosystem, research and development, and innovation output. It still needs to significantly improve its government effectiveness, regulatory environment, policy stability for business operations, infrastructure, open innovation, knowledge-intensive employment, and education.
If you like the newsletter, you will love to read our in-depth research and analysis at https://takshashila.org.in/high-tech-geopolitics.
Digital Disarmament 101
— Lokendra Sharma
Since it was established during the ninth Biological Weapons Convention (BWC) review conference, the Working Group on the Strengthening of the BWC has met four times and is scheduled to meet again in December 2024. The Pact for the Future, adopted by UN member-states during the Summit of the Future (22-23 September 2024), expressively calls for the ‘total elimination of nuclear weapons.’ The Pact further reaffirms that ‘[a]ny use of chemical and biological weapons by anyone, anywhere and under any circumstances is unacceptable.’ The Pact or the other outcome documents (Global Digital Compact and Declaration on Future Generations) of the Summit however do not mention one category of weapons at all — cyber weapons. On the cyber front, in fact, the progress of the international community has been abysmal. It took decades for the international community to finally agree to a draft UN Cybercrime convention in August this year. And we are nowhere close to a convention or arms control measures — similar to what we have for nuclear, biological and chemical — for cyber weapons.
But why does arms control not work in cyberspace? The following 2021 paper by Roguski has some of the answers:
Roguski, P. (2021). An Inspection Regime for Cyber Weapons: A Challenge Too Far? AJIL Unbound, 115, 111-115. DOI: https://doi.org/10.1017/aju.2021.6 (open access)
But first, what are cyber weapons? Building on the 2016 work of Trey Herr and Paul Rosenzweig, Roguski defines cyber weapons as ‘lines of code which affect the functioning of the target computer system. They can be conceptualised as essentially consisting of three elements: a propagation method, exploits, and a payload.’ Propagation method can be either direct (such as USB sticks) or indirect (such as through infected email or phishing) and refers to the way an access route is established to a target system. Exploit refers to the code that takes advantage of the vulnerabilities present in the target system to pave the way for the final component — payload — or even further propagation. Payload is the code that, when executed, fulfils the attackers’ goals.
Roguski argues that arms control measures cannot effectively target either of the three elements described above: propagation method, exploits, and a payload. Roguski argues that ‘propagation methods are dual-use, so any limitation regime could also potentially affect non-military uses of information and communication technologies.’ Arms control measures would not work for exploits for the same reason that prevents states from disclosing all vulnerabilities they discover. States choose to stockpile particular zero-day vulnerabilities for national security reasons.
While Roguski does not dedicate adequate attention to the difficulties of instituting measures for payloads specifically, he identifies two main problems that come in the way of arms control in cyberspace generally. First, ‘even if states manage to agree on what constitutes a cyber weapon and on which cyber weapons are to be subjected to a control regime, verification of compliance through inspections would pose the risk of compromising their function.’ This is because an inspection of a cyber weapon, according to Roguski, would entail an inspection of its code. However, cyber weapons would become useless once their code is known to other states, which can patch their vulnerabilities and deploy suitable detention and response mechanisms. And second, ‘on-site inspections would require access not only to particular sites, but to the entire governmental network, which no state would accept.’ This is because cyber weapons have intangible codes that ‘can be easily copied, stored, and transported.’ Both these factors differentiate cyber weapons from nuclear, chemical and biological weapons.
If not arms control, then what could work for cyber weapons? Roguski suggests a two-pronged approach involving ‘normative’ and ‘responsive-deterrent’ elements. While the ‘normative prong aims at the promotion of rules of responsible behavior in cyberspace’ through international law and international bodies, the ‘second prong aims at enforcing the rules of responsible behavior’ through sanctions and publicly attributing those behind cyber operations. However, Roguski does not suggest a concrete way of effectively operationalising both prongs, and this can potentially be a subject of further research.
What We're Reading (or Listening to)
[Opinion] A Partner in Taiwan, By Anushka Saxena
[Policy Paper] Environmental Cooperation: An Imperative for Subcontinental Thinking, by Adya Madhavan, Pranay Kotasthane, Y Nithiyanandam, Anand Arni and Nitin Pai
[Podcast] The Clock Is Tikking: Analysing TikTok's Addictive Algorithm, By Rohan Pai and Mariam Azeemuddin