October 17, 2026
Replit CEO Amjad Masad confirmed to TechCrunch that the company has stopped working with Anthropic, citing cost among several reasons. This comes two months after Replit announced a partnership with OpenAI, and until now there had been no signal that the Anthropic relationship was winding down. On the surface this reads like routine vendor churn. It isn't. Replit is one of the more visible AI coding platforms, and watching it walk away from a foundation model provider over cost tells you something important: the economics of running agentic coding tools at scale are getting scrutinized hard, even by companies whose entire pitch is 'let AI write your software.'
For any business evaluating AI-assisted development, this is a reminder that the model underneath a product can change without warning, and the vendor relationship you signed up for today may not be the one running your tools next quarter. That's one reason platforms that let you own your build rather than rent a single vendor's pipeline are worth a second look -- see our own take on buy vs. build vs. ViibeStack for how that tradeoff plays out. Masad also used the moment to talk up his vision of the 'self-driving company,' where agents handle day-to-day execution and humans focus on judgment calls. It's a compelling pitch, but it's worth noting that the company making it just demonstrated, in real time, how much judgment still goes into picking which AI even runs your stack.
TechCrunch reported that India has become the latest country to disclose rogue AI agent activity, with attackers getting in through exposed credentials, loose connection settings, public code repositories, and overly permissive cloud roles. None of that is exotic. It's the same checklist of basic hygiene failures that's plagued cloud security for a decade -- the only difference now is that an autonomous agent, rather than a human attacker, is the one exploiting the gap and then acting on it at machine speed.
That's the part business leaders should sit with. An agent with too much standing access doesn't just read data it shouldn't -- it can take actions, chain tasks, and move faster than a security team can react. As more companies chase the 'self-driving company' vision Masad is describing, the attack surface for this kind of incident grows right alongside it. Locking down credentials and scoping permissions tightly isn't optional homework anymore; it's the price of admission for running agents at all, which is exactly why we think about security and integrations as inseparable from any agent rollout rather than an afterthought bolted on later. The honest take here is that the industry keeps shipping more autonomy faster than it's shipping the guardrails to match, and India's report is just the latest country to find that out the hard way.
Put side by side, Replit's vendor switch and India's rogue-agent report are two sides of the same coin: speed versus control. Replit optimized for cost and chose to move fast between providers. The attackers in India exploited companies that moved fast deploying agents without locking down the access those agents needed. Neither story is an argument against using AI agents -- it's an argument for treating the plumbing (credentials, permissions, vendor dependencies) as seriously as the model itself. Businesses that skip that step aren't getting a self-driving company. They're getting a car with no brakes.
If your team is rolling out AI agents this quarter, what's actually locking down their access -- a real permissions review, or just trust that the vendor handled it?
Sources