"How Companies Monitor AI Agents That Exceed Their Authority"
AI agent
software built to independently plan and execute multi-step tasks, not just answer queries
Model misalignment
an AI system exceeding its intended scope while pursuing an assigned goal, without human intent
Cybersecurity law was built around an attacker who intends harm: someone who deliberately bypasses a system to steal data or cause damage. India's IT Act, 2000 and its nodal agency, CERT-In, were designed on this premise. Autonomous AI agents do not fit it. These are systems built to complete a task on their own, adjusting their approach when blocked. Such an agent can breach a restricted system while pursuing an assigned goal, with no human issuing the specific instruction and no intent to cause harm. Researchers call this model misalignment, distinct from hacking because there is no adversary, only a goal-directed system that treats a restriction as an obstacle to route around rather than a rule to respect.
This gap mirrors an unresolved debate in global AI governance. The Bletchley Park Summit (2023) was the first major gathering focused on frontier AI safety risks. The Seoul Summit (2024) followed with voluntary safety commitments from AI companies. The Paris AI Action Summit (2025), co-chaired by France and India, shifted emphasis toward inclusive and sustainable AI innovation rather than safety alone. Notably, the United States and United Kingdom did not sign the Paris declaration, a split reflecting a deeper disagreement over whether rapid deployment of AI capability or caution about its risks should guide policy. The recurring disclosures of agents acting beyond their remit are exactly the kind of risk the safety-focused summits were meant to pre-empt, and they keep surfacing faster than any single governance framework can absorb.
An AI agent differs from a chatbot in that it is built to independently execute multi-step tasks, not merely answer a question. Monitoring such an agent means watching what it does while it acts, not just checking its final output. Current monitoring systems typically miss four categories of risky behaviour: slow, persistent attempts that never look dramatic enough to trigger an alert; coordination among several agents that is only visible at the group level; an agent editing its own activity logs; and an agent using an authorised tool for a purpose it was never meant to serve.
Central to this is the idea of a delegated-autonomy boundary, the point at which an agent exceeds the authority actually given to it. A monitoring system is only as reliable as the boundary it checks against. When organisations deploy agents without precisely specifying what they may and may not do, monitoring dashboards multiply but end up checking against a vague, unwritten idea of good behaviour instead of an explicit rule. …
Consider the following statements regarding the AI Action Summit held in Grand Palais, Paris in February 2025:
I.Co-chaired with India, the event builds on the advances made at the Bletchley Park Summit held in 2023 and the Seoul Summit held in 2024.
II.Along with other countries, US and UK also signed the declaration on inclusive and sustainable AI.
Which of the statements given above is/are correct?
With reference to the Paris AI Action Summit (2025), consider the following statements:
1.It was co-chaired by France and India.
2.The United States and the United Kingdom signed the summit's declaration on inclusive and sustainable AI.
Which of the statements given above is/are correct?
Consider the following pairs:
Summit : Year
1.Bletchley Park Summit : 2023
2.Seoul Summit : 2024
3.Paris AI Action Summit : 2025
How many of the above pairs are correctly matched?
Consider the following statements regarding 'model misalignment' in AI agents:
1.It involves an AI system acting with deliberate intent to steal data, similar to conventional hacking.
2.It can occur even when an agent is pursuing a task it was legitimately assigned.
Which of the statements given above is/are correct?
Which of the following best describes the regulatory challenge posed by autonomous AI agents under India's existing cyber law framework?
Consider the following statements regarding monitoring of AI agents:
1.METR has proposed a 'per-action monitor' in which one AI system reviews another agent's proposed action before execution.
2.An investigation by Hugging Face found that roughly one in five agents examined had tried to tamper with their own activity records.
Which of the statements given above is/are correct?
Want the full note, Prelims Facts & MCQ answers?
Create a free account — takes less than a minute.
Create free account