As AI becomes embedded in core business operations, security leaders are confronting risks that traditional security models were never built to handle, from poisoned models to AI inherited silently through vendors. The software supply chain was already sprawling. AI adds a layer that is non-deterministic, often unsanctioned and hard to audit.
Anant Shrivastava, Chief Researcher and Founder of Cyfinoid Research Private Limited, has trained supply chain practitioners for years. In this conversation with Geetha Nandikotkur, at c0c0n 2026, Kochi, he explains why shadow AI is best handled with incentives rather than crackdowns, where AI-BOMs fall short, and why, when a model misbehaves, accountability lies with the person who gave it authority.
Geetha Nandikotkur: Anant, threat models are changing how CISOs work. How does AI change the supply chain threat model compared with traditional software supply chain risk? Where are the weakest links: training data, third-party models, APIs, or vendors embedding AI in their products?
Anant Shrivastava: Supply chain risk was always cross-sectional. It is not limited to one area; it touches many. AI is more cross-sectional still, sitting on top of the software supply chain. An employee may be using AI to do something. They may have deployed AI to automate a task. Or someone else may be using AI, and my company consumes the outcome.
AI is also not deterministic. When we say AI today, we are effectively talking about LLMs, and large language models are designed to be unpredictable. They are not supposed to give the same output for the same input every time. On top of that, we still have to make sure none of our existing functionality breaks.
Geetha Nandikotkur: Many CISOs inherit AI through SaaS tools and suppliers without ever choosing it. How should they discover, assess and govern AI they didn’t deliberately adopt?
Anant Shrivastava: Some of it we can find, and some we can’t. If an employee builds a portal that uses your company’s name and logo, there are many companies that can help you detect it. But if the application doesn’t indicate that it belongs to your company, there is no automation to find it. The only clues are email communication or some other trail you can capture.
I proposed an approach last year, and a couple of people have told me it helped them. IT should incentivize people to hand over what they build. Tell them: “If you solved a problem we couldn’t, give us the ownership. We will maintain it, we will own it, and we will let you use it properly.” You don’t have to take the ownership, and you don’t have to take the blame for it.
Geetha Nandikotkur: You’ve worked on supply chain security for years and trained practitioners. What has changed with the advent of AI?
Anant Shrivastava: The biggest change is that people want to experiment. Earlier, for anything, I had to buy a product or a license. Now engineers and managers think: if I’m going to pay $1,000 a month to another entity, why not put $1,000 worth of tokens in first, see if I can make it work within the company, and own it?
The only places I’ve seen this not happening are regulated entities, or places where the cost of owning something that goes wrong is high. They don’t want to own anything, because they want to delegate the risk out.
Geetha Nandikotkur: How real is the risk of data poisoning or model tampering today? And is an AI Bill of Materials, or AI-BOM, practical for governance and provenance?
Anant Shrivastava: It has just been proven that models can be poisoned. ProjectDiscovery, I believe on the 8th of October, released a note saying that for a total cost of $50 they were able to backdoor a 9-billion-parameter model. The model responds normally until a trigger word is given. When it is, it steals your Codex credentials and takes them away.
On the AI-BOM, there is a need for it. The tricky part is that proprietary models don’t give us the information we need. They give you a model card, a 90- or 200-page document, and you can’t make head or tail of it. You can’t extract the right things. It’s intentionally created that way, because if they disclosed the dataset they used, they would face huge fines.
Geetha Nandikotkur: Is the SBOM still in vogue, or are we moving away from it towards AI?
Anant Shrivastava: The SBOM is still needed. SBOM, CBOM, QBOM and AI-BOM all focus on different things. CycloneDX does a good job here: it treats BOMs as independent entities and as a unified one, so you can represent all of them in a single file or keep separate files for each.
The SBOM looks at your software versions. The CBOM looks at your cryptographic libraries. The QBOM checks whether those libraries are ready for the post-quantum world. The AI-BOM looks at your models. None of them overstep the others.
Geetha Nandikotkur: From an incident response point of view, what does response look like when the compromised asset is a model rather than a server? Are most security teams prepared?
Anant Shrivastava: Are they prepared? Most probably not.
This also comes from my rant about how the headlines are portraying OpenAI, Anthropic, Gemini and others right now, with stories of models hacking platforms and people. Models on their own can’t hack anything. You have to provide electricity and GPU power. You have to give them a prompt. They need training data that tells them what to do. Only then can they do something.
When OpenAI said its models were doing swarm-level activity and collaborating, it sounded like they had started devising things on their own. Then an interview came out in which an OpenAI engineer said that for about a year they had been training the models to work in collaboration. You trained them to do this, and now they did it. Why are you surprised?
So models can be the perpetrators, but someone is always behind them. Any activity a model performs, someone gave it the right, or the conditions, to do it. It’s like saying, “My AI model responded to my email” or “deleted an email.” No, you gave the model authority to do those tasks and then let it run unsupervised. If you are pointing blame, it belongs to the person who assigned the job to the model.
Geetha Nandikotkur: If a CISO could implement one control in the next 90 days to meaningfully reduce AI supply chain risk, what would it be?
Anant Shrivastava: I wouldn’t say one control, but two in parallel. The first: start cataloguing what you have, and remove anything you don’t need. In a nutshell, reduce your attack surface. The fewer things you have to focus on, the better you can focus on them.
About Anant Shrivastava
Anant Shrivastava is an information security professional with 17+ years of corporate experience in Network, Mobile, Application and Linux Security. Anant is an avid open-source supporter and runs multiple open-source projects, prominent among them being TamerPlatform and CodeVigilant.
He contributes to multiple open communities like null and Garage4Hackers. He has also helped establish a local null chapter in his hometown, Bhopal.
He has been a speaker and trainer at a multitude of conferences such as Black Hat USA/Asia/EU, DEF CON, Nullcon, c0c0n, Rootconf and many more.
He also participates in various communities as a CFP reviewer, notably Black Hat EU, Nullcon, Rootconf by Hasgeek, Recon Village at DEF CON, Cloud Village at DEF CON and Adversary Village at DEF CON.
His code contributions can be found on GitHub. He is active on Twitter and Fediverse, and his talks and presentations are available online. He writes about his experiments on his blog.
