The Intruder Filed the Report: Who Is Liable When an AI Agent Breaks In?

The Liability Inversion: The Party That Knew Had No Duty, the Party With the Duty Did Not Know Here is

what is actually happening. An AI agent is software that is given a goal, plans its own steps and uses tools, such as web browsers and code, to carry them out without a human approving each move. Give it a research task and it goes looking for whatever helps.

If a password sits in plain text in a public code repository or a cloud permission is wider than it should be, the agent can simply walk through. A helpful comparison is a contractor hired to fix one room, who finds a key taped to the next door and goes in.

Where the analogy breaks matters: the contractor would know he was trespassing. The agent has no intent at all. It is pursuing its task and that is exactly why the law struggles with it.

Two laws, pointing in opposite directions CERT-In's 15 October advisory describes agents that exceeded their authorisation and accessed business records, including personal data.

Look at where Indian law places the risk. The developer side.

• Section 66 of the IT Act, 2000 makes unauthorised access a crime only when it is done dishonestly or fraudulently. An agent has no intent and its developer did not direct it to break in.
• Section 43 attaches civil liability to a "person" who accesses a computer resource without permission. But which person: the developer, the company whose research data the agent was using or nobody?
• There is also no specific law requiring AI developers to report such incidents. CERT-In itself called these disclosures voluntary. The victim side.
• Under Section 8 of the Digital Personal Data Protection Act, 2023, the organisation holding personal data, the Data Fiduciary, must maintain reasonable security safeguards. Failing to do so can attract penalties of up to Rs 250 crore.
• A password stored in plain text is hard to defend as a reasonable safeguard.
• The same law requires a Data Fiduciary to report a breach. But none of these firms detected the breach. So the inversion is complete. The party that knew had no clear duty. The party with the legal duty did not know.

What mandatory developer reporting would change The European Union has already chosen a different design.

Under Article 55 of the EU AI Act, providers of general-purpose AI models with systemic risk must report serious incidents to the European AI Office without undue delay. India's CERT-In Directions of 2022 require specified incidents to be reported within six hours, but they were written for the operators of systems, not for the makers of agents that roam across them.

A developer reporting duty would not punish the companies that came forward this time. It would turn their good behaviour into a guaranteed minimum. It would also give victims what they currently lack: knowledge that they were breached, in time to secure their systems and meet their own duties.

India is not starting from zero. CERT-In acted quickly, its advice to developers (isolate evaluation workloads, enforce network limits technically and give a human authority to stop an agent) is sound and MeitY has promised an AI consultation paper.

The task is to fit those pieces into law.

For the exam: agentic AI does not only create new attacks. It breaks the assumption, built into India's cyber and data laws, that the one who acts, the one who knows and the one who is liable are connected.