Hi Friends,

Even as I launch this today ( my 80th Birthday ), I realize that there is yet so much to say and do. There is just no time to look back, no time to wonder,"Will anyone read these pages?"

With regards,
Hemen Parekh
27 June 2013

Now as I approach my 90th birthday ( 27 June 2023 ) , I invite you to visit my Digital Avatar ( www.hemenparekh.ai ) – and continue chatting with me , even when I am no more here physically

Translate

Wednesday, 18 November 2026

92% hit by an AI agent incident, India clears 1 of my 7 controls

92% hit by an AI agent incident, India clears 1 of my 7 controls

ServiceNow's survey of 500 Indian security leaders finds 92% of organisations had an AI agent security incident in the past year, and only 36% contained it within 24 hours. Measured against the seven controls I published in March, India clears one - and clears it by accident. I suggest three fixes before CERT-In's October advisory turns into a filing exercise.

On 18 November, in New Delhi, ServiceNow released Deployed but Ungoverned: The State of Agentic Security in India, a survey of 500 security leaders. Kulmeet Bawa, who heads India and SAARC for the company, put the whole difficulty in one sentence: "You can't secure or govern what you can't see."

Here is what the 500 told them (ServiceNow, 18 November 2026):

  • 92% had an agentic security incident in the past year; only 36% of them contained it within 24 hours
  • 19% have a single named individual accountable for agentic AI security; 18% have unified, single-interface visibility into their agents
  • 37% discovered agents running in production that were never registered or known to security; only 30% enforce an automated deployment gate
  • 51% have automated real-time detection of anomalous agent behaviour, but 67% cannot act on it automatically
  • 83% run agents in or touching operational technology or physical environments; only 29% have a governance framework for them

Amar Ujala carried it in this morning's Hindi business diary, and Inc42's report the same day quotes Palo Alto's Nikesh Arora saying AI is "making it easy for hackers to attack multiple customers." The week has a mood. The survey has numbers.

In March I wrote When AI Gets Keys, after reading the Times of India report on the Agents of Chaos experiments, and listed seven controls that should be built in rather than bolted on. One of them:

"Least privilege by design: agents should only have the narrowest tool access they truly need. No default keys to mailboxes, shells or admin controls."

I have waited eight months to be proved wrong. Here is my own scorecard, with the nearest survey question set against each control:

My control, March 2026Survey finding, 18 Nov 2026In place?
Least privilege by design37% found unregistered agents in production; 30% gate deploymentNo
Human approval before irreversible actions67% cannot act automatically on an anomalyYes, by accident
Immutable state checks and audit trail47% can produce a complete evidence packageNo

The other four - purpose-bound tokens, cryptographic identity binding, whitelisted rule sources, a private deliberation surface - the questionnaire does not ask about at all. When a survey of 500 security leaders has no question for a control, that control is not yet a market requirement.

So: one out of seven, and the one is no compliment. The human is still pressing the button because nobody has built the automatic brake. The report calls this a response gap. I would call it a kill switch that happens to be a person, with none of the immutable logging my third control asks for.

Meanwhile CERT-In has entered the room. Its October 2026 advisory on AI agents, CIAD-2026-0049, puts agent visibility and control on the agenda, and under its 2022 directions an Indian organisation must report a cyber incident within six hours of becoming aware of it. Read that beside 19% named accountability and 18% visibility and the real problem appears: you cannot report in six hours an agent you cannot name.

Three things I would ask for, in order of usefulness:

  1. A registration record and a named human owner filed before any agent goes live - no agent in production that security has not seen.
  2. A brake a security tool can pull, not only one a human can reach. The 67% is a kill-switch statistic before it is a response statistic.
  3. The six-hour filing generated from an evidence package already stored, rather than assembled by hand after midnight.

The advisory was the easy part. The hard part is that 19 out of 100 boards can tell you whose phone rings at three in the morning. Until that number moves, 92% is not a statistic about attackers. It is a statistic about us.

With regards,
Hemen Parekh
www.Iam-Immortal.ai / www.IndiaAGI.ai / www.YourContentCreator.in
19 Nov 2026

No comments:

Post a Comment