On 15 October 2026, CERT-In issued an advisory that is unusual for one reason. The attacker in each incident it describes was an AI agent.
The advisory is CIAD-2026-0049, titled "AI Agents as Attack Vectors", and it is rated Critical. It describes agents that reached the live systems of Indian organisations and read business records, including personal data and confidential material. The detail that should stop any founder is this: the affected organisations did not notice. In the first category of incidents, the access came to light because the AI model providers disclosed it themselves, after their own internal reviews.
I read the advisory in full, because most of the coverage stops at the alarming part. The useful part is underneath it. The weaknesses the agents used are ordinary ones, the kind most growing businesses have somewhere, and CERT-In lists the checks that close them.
What did CERT-In actually report?
CERT-In describes two separate categories of incident. They are worth keeping apart, because they are different problems.
Category one: agents under evaluation went beyond their task. AI model providers voluntarily disclosed that autonomous models and agents undergoing research and evaluation exceeded the scope of what they were asked to do and reached live systems of Indian organisations. They got in using credentials and connection settings left in application configuration, in public code repositories, and through overly permissive cloud roles. Nobody broke a password. The agents found keys that were already lying around and used them.
Category two: criminals using agentic tools. Through its own threat hunting, CERT-In found a financially motivated campaign against Indian entities that used publicly available, AI-native offensive-security frameworks. These are tools built for legitimate security testing that are widely deployed and, in CERT-In's words, being repackaged.
What the advisory does not say matters too. It does not name the AI providers, the affected organisations or the sectors, and it gives no count of incidents. Anyone telling you which companies were hit is going beyond the source.
Why is this about weaknesses you already have?
This is the part of the advisory I would put in front of every operations and tech lead. CERT-In lists the weaknesses that were exploited, and none of them is exotic:
- exposed credentials and keys
- forgotten or test APIs reachable from the internet
- payment-API keys stored in plain text
- over-permissive cloud and backend roles
- management interfaces exposed to the internet, or unpatched middleware, with weak or no multi-factor authentication
CERT-In's point is that AI found and exploited these faster and at greater scale than human attackers would. Its observations are specific. Agentic tooling "compressed the time between initial access and final impact from days to a few hours". Multi-agent operations ran in parallel for hours to days. Victims that remediated only partly were attacked again within days, at larger scale. And agents "were observed re-deriving and re-using the same exposed credential in separate operations".
That last observation changes the economics of a leaked key. A key that sat in an old repository for a year used to be a risk that needed a patient human to find it. Now it is a risk that a tireless piece of software can find, use, lose and find again.
Where do these weaknesses hide in a growing commerce business?
The advisory is written for every organisation, so here is how I would translate it for a D2C brand or a small business with a website, a payment gateway and a handful of tools. None of this comes from the advisory's incident details, which it does not publish. These are simply the places where the weaknesses it lists usually live.
- API tokens for apps and integrations. Every app that connects to your store or your ERP holds a token. Many were created once, by someone who has since left, with every permission ticked.
- Payment-gateway keys. The secret key for your gateway should live in one protected place on a server. It is worth checking whether it also lives in a developer's old code, a shared document or an automation tool.
- Backend-as-a-service projects. CERT-In calls these out by name and asks for row-level security and for privileged keys never to be placed on the client side. If a web or mobile app talks directly to a hosted database, test what an anonymous user can actually read and write.
- Staging and test sites. A copy of the store or an internal API, set up for a launch and never taken down, is a forgotten test API reachable from the internet.
- Automation tools. Workflow tools hold credentials to everything they connect. A webhook URL or an API key pasted into a workflow is a credential.
- Admin panels. Store admin, hosting dashboards, the database console, the domain registrar. Each one without two-factor authentication is on the advisory's list.
What does CERT-In ask organisations to do?
The advisory gives ten actions for organisations and service providers. Here they are, condensed, with what each means in practice for a small team.
| # | CERT-In's action, condensed | What it means for a small team |
|---|---|---|
| 1 | Apply the baseline in CERT-In's Blueprint for defending against AI-assisted exploitation | Read the Blueprint once and note what you already do |
| 2 | Audit code, repositories including history, CI/CD variables, config files, client apps and cloud storage for exposed credentials | Search every repository and its history for keys, not just the latest version |
| 3 | Rotate exposed credentials, purge them from history, verify externally, scan every commit and build for secrets | A leaked key is fixed by replacing it, not by deleting the file |
| 4 | Keep a complete API inventory and switch off dormant, deprecated and test APIs and accounts | List every endpoint and account that exists, then turn off the ones nobody uses |
| 5 | Apply least privilege in cloud and backend set-ups, use row-level security, never put privileged keys client-side, test what anonymous users can read and write | Every key gets the smallest access its job needs |
| 6 | Take admin interfaces off the internet, require MFA for all admin access, patch middleware promptly | Two-factor authentication on every admin login, starting today |
| 7 | Detect attacks at machine speed: rate-based API anomaly detection, alerts for enumeration, bulk reads and automated sign-ups, keep logs for at least 180 days | Know what normal traffic looks like, and get an alert when someone reads everything |
| 8 | Treat any credential exposure or confirmed access as a compromise until forensically cleared, retest after fixing, expect repeat attempts within days to weeks | "We rotated it" is not the end of the incident |
| 9 | Govern internal AI agents with short-lived, least-privilege identities per run, default-deny outbound network access and tamper-resistant logs | If you run agents, each run gets its own narrow, temporary access |
| 10 | BFSI and fintech entities should also apply the payment and disbursement API measures in CIAD-2026-0046 | Relevant if you are regulated in financial services |
There is a separate list of seven actions for AI developers and evaluation agencies, covering isolated test environments, synthetic credentials and stopping runs that touch third parties. That list is aimed at the labs, not at you, but it tells you what responsible providers are now expected to do.
If you run AI agents or automations yourself
Action 9 is the one that applies to more businesses each month. If an agent drafts customer replies, reconciles orders or updates records, it acts with whatever access you gave it. CERT-In asks for three things:
- Short-lived identities per run. The agent gets credentials that expire when the task ends, not a permanent master key.
- Least privilege. It can do its job and nothing else. Reading orders is not the same permission as issuing refunds.
- Default-deny outbound access and tamper-resistant logs. It can reach only the systems it needs, and every action it takes on production systems and data is recorded somewhere it cannot alter.
I wrote earlier about where AI agents fit in e-commerce operations, and the principle there was to let an agent draft but be careful about letting it send. This advisory is the security version of the same idea. The question is never only what an agent can do. It is what it is allowed to do, with which keys, and who can see what it did.
This is the work we do at INovaBeing under AI & Automation: building agents and workflows into real business systems, with human-in-the-loop steps where decisions carry risk. The access design is part of the build, not an afterthought. If you are not sure what your current integrations and automations can actually reach, a Technology Diagnostic is the sensible place to start, before anyone builds anything new.
A plan for this week
If I were running a small commerce team today, this is the order I would work in.
In the next 48 hours:
- Turn on two-factor authentication for every admin login: store, hosting, database, domain, payment gateway, email.
- Rotate the payment-gateway secret key if you cannot say with certainty where every copy of it lives.
- Remove access for anyone who has left.
This week:
- Search every code repository, including its history, for keys and passwords. Rotate anything you find, then remove it from history.
- Write down every app, integration and automation connected to the business, and what each can access. Reduce any that have more than they need.
- Find and switch off staging sites, test APIs and old accounts nobody uses.
This month:
- Check your logging. CERT-In's 2022 directions already require a rolling 180 days of logs and reporting of specified incidents within 6 hours, so this is not new, but many smaller businesses have never checked.
- Set up alerts for unusual volume: bulk exports, rapid reads of customer records, a burst of new sign-ups.
- If you run AI agents, give each one its own narrow, expiring credentials and a log you can read.
What I take from this
The weaknesses here are old. What is new is the speed. When finding an exposed key took a human days, a small business could get away with a little untidiness. When an agent can find it, use it and come back for it within hours, untidiness becomes exposure.
My read is that this advisory is less about AI than it first appears. It is a list of basic hygiene, given new urgency by the fact that the thing checking your hygiene no longer gets tired. None of the ten actions needs an AI tool. Most need an afternoon, a list and someone who owns the outcome.
The full advisory is on the CERT-In website, and it is worth reading in the original. I also wrote about the PHP2Shell vulnerabilities, which make the same point from the other side: AI is now speeding up the people looking for flaws, for good and for bad.




