On 6 November I turned CERT-In's AI audit guidelines into four things an auditor could actually demand, and said the six-hour incident clock was unenforceable without them. Three weeks later, a ServiceNow and Phronesis survey of 500 Indian security leaders supplies the failure rates behind that claim: 92% had an agent incident, only 36% contained it in a day, 19% have one named owner. This post reads CIAD-2026-0049 as compliance guidance, not commentary — a permission ledger, an Agent Passport with a tested kill-switch, prompt-injection red-teaming in every audit, and a public register of agent incidents.
On 6 November, in CERT-In just wrote the audit rules I asked for in "When AI Gets Keys", I reduced the guidelines to four artefacts an auditor could demand, and then wrote a line I had no evidence for: you cannot report an agent incident at six hours if you never logged it. This week the evidence arrived, and it is worse than I assumed.
The Times of India reported on 19 November 2026 on a ServiceNow study, carried out with Phronesis Partners, titled "Deployed but Ungoverned: The State of Agentic Security in India" — 500 Indian security leaders inside a global sample of 3,700 across 20 countries. The numbers land almost one for one on my four asks:
- 92% had at least one agentic security incident in the past year; only 36% contained it within 24 hours.
- 19% have a single named individual accountable for agentic AI governance; 18% have single-interface visibility across their agents.
- 30% enforce an automated gate that blocks an agent until it passes security checks; 47% can produce a complete evidence package on demand.
- 51% detect anomalous agent behaviour automatically; 67% cannot automatically act on it.
- 37% found agents in production in the last 12 months that security had never registered.
- 83% run agents in or alongside operational technology; 29% have a governance framework built for it.
"You can't secure or govern what you can't see. Every agent should be registered, owned and visible in one place." — Kulmeet Bawa, India and SAARC, ServiceNow, 19 November 2026.
Half the leaders surveyed called their agentic security "established". That is the sentence I would staple to the top of the file, because confidence and control have come apart.
Read as compliance guidance under CERT-In's October 2026 advisory on AI agents, CIAD-2026-0049, the study stops being commentary and becomes a list of failure conditions. An auditor can now ask for four things and fail an absence:
- The permission ledger answers 18% and 37% — one view, every agent, who granted its reach, when it lapses.
- The Agent Passport answers 19% and 30% — scope, blast radius, a named human owner, a kill-switch carrying a test date.
- The prompt-injection red-team clause answers 67%. Detection without action is a siren, not a control.
- The public incident register answers 36% and 47%. Containment in a day is learned, not discovered.
The operational-technology figure should stop a board. An agent that can open a valve is not an IT asset with a login, and Express Computer's report of 19 November 2026 makes the gap plain: 83% are running there, 29% have a framework for it.
Liability has also stopped waiting for guidance. On 18 November 2026 Arizona's attorney general, Kris Mayes, said her office had opened a criminal investigation into OpenAI and other AI companies over the "apparent infiltration of banks by rogue AI agents", and is examining whether state law is being violated on an ongoing basis; a criminal investigation does not mean charges follow. The Product Growth brief of 19 November flagged it. I would not treat it as American noise.
None of the four artefacts is a technology purchase. A one-agent startup can keep its passport to a page; a bank cannot. In "When AI Gets Keys" (24 March 2026) I asked for containment standards and auditability when there was no survey to quote. There is one now, and it says the principles are still unbuilt.
So, to the Director General of CERT-In: the six-hour clock is only as good as the ledger behind it. Make the Agent Passport the unit of compliance, with a named owner and a kill-switch that carries a test date, and the clock becomes a clock instead of a wish.
With regards,
Hemen Parekh
www.Iam-Immortal.ai / www.IndiaAGI.ai / www.YourContentCreator.in
20 Nov 2026
No comments:
Post a Comment