Opinion
Seizing the opportunities of vibe coding without paying the price
“The policy should be identical for humans and for agents: the same rules, the same transparency, the same level of supervision. Otherwise, ‘blind spots’ are created where innovation moves faster than protection,” writes Itai Schwartz, CTO and co-founder of cybersecurity startup, MIND Security.
There’s a feeling that the world speaking in vibe coding is running faster than all of us. Instead of sitting and writing code line by line, we tell the system what we want, and it generates it for us at incredible speed. On the business side, it’s a dream: fewer bottlenecks, more versions, shorter time to market, and small organizations that suddenly look like huge development divisions.
But this dream comes with a price tag that doesn’t appear in the presentation: the data.
To write “smart” code, these tools need to touch keys, access tokens, internal documents, and customer data. When they work correctly - we win. When they fail, the mistake is measured not only in broken lines of code, but in trust, reputation, and company value.
The common mental mistake is thinking that leaks are always the fault of a careless developer. In the age of AI agents, the “personality” performing the work is a process-oriented entity whose goal is to complete a task - not to assess risks. If it’s convenient for it to pull an example from the CRM to build a test, it will do that. If an access token is found in its memory and seems relevant, it might slip into code or documentation without any malicious intent.
That’s how what I call context contamination is created: information born in one place ends up in another - tests, comments, product documents - simply because it was convenient for the machine. And this problem is worsened when advanced protocols give organized but very broad access to enterprise systems. Each such connection shortens the way to a solution but also opens a possibility for a small mistake to become a big event.
Yes, in the past year there have been several cases that made headlines, including around Base44, showing how quickly a “small flaw” in connection or redirection can turn into an account takeover or data exposure. It’s important to understand: this is not a phenomenon limited to one platform, but a symptom of an era. When the same infrastructure hosts many applications and many users, the boundary between code, identities, and data becomes thinner than ever. No malicious intent is needed - just an inappropriate authentication design, a wrong redirect, or an overly generous API access point can lead to the exposure of access keys.
And when that happens, an organization doesn’t deal with a “bug” - but with a trust event.
A customer who finds their personal data in a test file will not forget it; an investor who reads about a leak will question how the company controls its data; and the regulator will not overlook fines.
The natural reaction when talking about protection is “let’s review the code better.” That’s important - but it comes too late. If the data has already entered the system, the damage may already have been done. In my opinion, the right way is to focus protection on the moment when the data actually moves - when the agent accesses it, uses it, and returns an answer. That’s the moment to stop the exposure of secrets in real time; to ensure that real information does not leave its dedicated environment; to separate tasks so that what is done in one doesn’t leak to another; and to give each access limited and temporary permissions only.
The most important principle is that the policy should be identical for humans and for agents: the same rules, the same transparency, the same level of supervision. Otherwise, “blind spots” are created where innovation moves faster than protection.
For this to work, organizations need to see data as an asset that crosses teams and tools - not just a “byproduct of development.”
That means adding data metrics to management dashboards, measuring leaks like build times, and understanding that the cost of a data incident is not only technical but also business-related: delayed deals, harm to company value, regulatory risk.
Those who manage the flows will enjoy the speed of vibe coding - without paying compound interest.
Within this space, quite a few companies are emerging that try to place protection exactly where it’s needed - between the agent and the systems.
A middle layer that sees the movement in real time and stops leaks as they occur, with uniform enforcement for both people and agents.
Such solutions make it possible to solve a difficult problem elegantly and return control over data to the organization without slowing down the pace that vibe coding promised.
Itai Schwartz is the Co-Founder and CTO of the cybersecurity startup MIND Security.
