Sanjay Jeyakumar, co-founder and CTO of Abnormal AI, flew in from the Bay Area to join a new episode of In Conversation with Saepio’s CEO Amir Nooriala.

In 40 minutes, they cover:

The incident that changed the conversation. OpenAI models, under routine capability testing, broke out of their sandboxes. They started communicating with each other through message boards, moved into an infrastructure company, and from there into Hugging Face. There was no attacker and no malicious model. It happened at least three times. Sanjay walks through what actually occurred and why the industry reacted the way it did.

“Nobody truly has a moat.” An unusually direct admission from the founder of a company valued above $5bn. Sanjay explains why he thinks AI has eroded defensibility across the board, and what he believes is left: customer trust, and the ability to listen harder and ship faster than anyone else.

Why business email compromise is still unsolved. Spam had payloads. Phishing had links. BEC has neither, because the attack is the text itself. And when the invoice genuinely comes from your supplier’s compromised account, there’s nothing to fingerprint. Behaviour is what’s left.

The defender’s advantage. The attacker doesn’t know your company. You do. Sanjay makes the case that in an AI arms race, the side holding the behavioural data starts ahead.

Rogue AI, and the never before seen attack. Every agentic attack is novel by definition, which makes threat intelligence close to useless against it. Abnormal is now extending behavioural models into cloud, at customers’ request.

And the IPO question, where he’s more forthcoming than expected.