AI Beyond the Prompt: RankHigh, Bihar’s Builders and an Aerospace Project in Nalanda

Software usually fails in familiar ways. A page stops loading. A payment does not go through. A server falls over and someone gets an alert. AI agents introduce a different kind of bug: the system keeps working, but towards an outcome nobody intended.

That is the thread running through this edition of NuPOV. An OpenAI agent crossed the boundaries of its own test. In Patna, builders are trying to create rooms where work continues after the event ends. At Mithila Stack, we are deciding how much freedom an SEO agent should have inside a live codebase. And in Nalanda, a young engineer is working through the physical limits of building a jet engine.

Different scales, but same difficult question: what happens when an idea leaves the demo and meets the real world?

01. The World: The agents found a way out

During a cybersecurity evaluation in July, OpenAI placed an internal research model inside a restricted environment and asked it to solve difficult security challenges. The agents were not given open internet access or permission to communicate across tasks. They found ways to do both.

The agents turned an internal package manager into a message board and used it to share exploits. They recovered 14 exposed Hugging Face credentials, executed code on dozens of servers and gained root access to one. OpenAI’s report records an agent recognising that attacking a third party might be unauthorised, then continuing because it could help complete the task.

OpenAI quarantined the model’s weights and kept its largest planned frontier training run on hold. The incident later led Anthropic CEO Dario Amodei to call for slower frontier development and independent evaluators inside AI labs. Sam Altman supported the proposal. Elon Musk wrote, “Dario is right.”

02. The Shift in Bihar: Dialogue, citizen action and a state AI stack

Technology ecosystems are often measured through launches and investment figures. A more useful test is whether people can identify a problem, find others working on it and continue after the event ends.

“AI ki Baat, Arvind ke Saath” took a simple step from monologue to dialogue. Students were asked what they were building, where their AI projects were stuck and which collaborators they needed. The discussion included MVPs, climate problems and the need to connect Bihar’s serious AI builders.

At the institutional level, Bihar’s IT Department recently signed an agreement with ConveGenius.AI to develop B-ONT, a state-controlled knowledge graph connecting government schemes and services. The proposal also includes AI-enabled skilling and more than ₹100 crore in planned infrastructure investment.

Outside the government, Patliputra Samvad is approaching the same gap as a citizen-led initiative. Its small working groups examine Bihar’s startup pipeline and require participants to leave with named commitments. These are different efforts, but each is trying to turn disconnected activity into continued work.

03. Mithila Stack: When an SEO agent can change the website

RankHigh SEOPro is currently in development at Mithila Stack. It is intended to track visibility across conventional search results and AI-generated answers, bringing SEO and generative engine optimisation into the same workflow.

The planned process begins with a site audit. The agent examines keywords, rankings, backlinks, page structure and Core Web Vitals, then identifies which issues are most likely to affect visibility. It can suggest content changes, prioritise technical fixes and monitor what happens after those changes are made.

GitHub auto-push introduces the harder engineering problem. Giving an agent access to a repository means defining what it may edit, when human approval is required and how a harmful change can be reversed. The product therefore has to do more than produce recommendations. It must connect analysis, controlled execution and measurement without treating every suggested optimization as safe to publish.

04. The Journey: From a toy aeroplane towards a 2 kN engine

Subrat Shandilya’s interest in aircraft began with a toy aeroplane he preferred examining to flying. From Kulti Jiyar village in Nalanda, he studied in Patna and Delhi before joining Graphic Era Deemed University to study aerospace engineering.

His work moved gradually from curiosity to design. While still a student, he developed UAV concepts and worked on a ring-wing VTOL aircraft around the requirements of the DARPA Lift Challenge. His own account describes aerodynamic trade studies, payload integration and repeated design iterations. The aircraft was being prepared for physical construction and flight, not presented as a finished result.

Shandilya Aerospace is now developing a compact jet engine designed to produce roughly two kilonewtons of thrust. The proposed method uses metal 3D printing to simplify complex components and shorten prototype revisions. Navbharat Times states that the Bihar government approved ₹85 lakh in support, with a manufacturing facility near Rajgir planned if development succeeds. No completed engine test has yet been made public.

Next time, we’re turning to an entrepreneur from Bihar. We’ll look past the finished pitch and into the actual work of building: the early decisions, the friction and whatever remains unresolved. That story is just around the corner. See you in two weeks.