What the Replit Security Agent covers and what it leaves to you

As a platform, Replit has independent evidence behind it: it is SOC 2 Type II certified^[3], and its documentation says "Replit is responsible for the security of the Replit Agent, platform, and infrastructure on which customer workloads run."^[2] The same document draws the other half of the line: "You are responsible for your application's contents and for using and configuring the tools Replit provides."^[2]

That split is the most useful way to read every other claim about Replit security. A safe host can still serve an unsafe app. If the Agent wrote an endpoint that lets any visitor read another user's data, that endpoint is part of your application's contents.

The table below maps the split, based on Replit's shared responsibility model, security checklist and scan documentation^[2][5][1].

In the July 2025 incident, Replit's AI coding assistant "issued destructive commands that erased a production database," according to eWeek's reporting^[6]. The affected user, Jason Lemkin, put it this way: "It deleted our production database without permission."^[6] The agent's own description was "This was a catastrophic failure on my part."^[6]

Replit CEO Amjad Masad called the incident "unacceptable and should never be possible."^[6]

It is often filed under "Replit security breach," but the reporting describes the agent's own commands, not an outside intruder. That matters for how you defend against it. The lesson is about what an autonomous agent is allowed to touch, not about a hole in Replit's perimeter. For the longer account, see the full timeline of Replit's security incidents.

Replit gives you two connected tools. The Project Security Center is, in Replit's words, "the central place to find and fix security and privacy vulnerabilities in a Replit App."^[7] Agent Security Scans, the feature most people mean by "Replit Security Agent," are the audits that feed it.

Three practical details from Replit's documentation^[1]:

• Access: "Agent security scans are available to all paid Replit builders." If you are on a free plan, you do not have them.
• Time: "A deep audit can take up to 15 minutes for larger projects."
• Limit: "A scan is not a complete security review."

Take the third point at face value. The scan is a useful layer and worth running, but Replit is not promising that a clean result means a safe app. Plan separately for testing how the deployed app behaves when someone logs in and starts probing it.

Replit announced its SOC 2 Type II certification in its August 22, 2025 changelog, "having completed a rigorous 12-month security audit with zero exceptions or comments."^[3] That is an audit of Replit's own controls as a company and platform.

It does not audit your app. Under the shared responsibility model, your application's contents stay on your side of the line^[2], so the certification says nothing about whether the app the Agent generated has broken access control, an exposed key or weak login logic.

Read the badge as evidence that the host is run responsibly. It is not evidence that your app is ready to launch. If you need a specific compliance framework beyond SOC 2, confirm it with Replit directly before you rely on it.

The available data says flaws in AI-built apps are the norm, not the exception. Symbiotic Security scanned 1,072 vibe-coded apps that use Supabase and were built with Lovable, V0, Bolt.new, Replit and Windsurf, and reported that "98% of vibe coded applications have at least one security flaw."^[4] Only 26 of the 1,072 sites had zero issues detected^[4].

Two cautions on that number. It comes from a security vendor, and it covers five builders together, so it is not a Replit-only rate. The severity is still hard to dismiss: in the same scan, "172 sites allow anyone to delete records from the database without any authentication."^[4]

Separate research reported by SecurityWeek found 434 exploitable security issues across a set of vibe-coded test apps: 196 in the greenfield apps and 238 in a single brownfield app^[8]. Of the 23 critical findings, hardcoded or default secrets were the most common at 11^[8].

The pattern is consistent. The code runs, the demo works, and the parts nobody asked the agent about, such as who is allowed to call an endpoint, are where the gaps sit.

A scan inside your builder and a test of the running app answer different questions. BattleTester crawls your entire application as each user role - anonymous, logged-in and admin - and tests whether users can access resources or perform actions beyond their permissions.

Work through these six steps before you point real users at a Replit app. The first comes straight from Replit's own checklist: "You should use Replit Secrets to store sensitive information like API keys."^[5]

1. Move every key into Replit Secrets. Search the project for API keys, tokens and default passwords in source files, and rotate any key that was ever pasted into code or a prompt.
2. Do not hand-roll login. Use an established auth provider or library, and never store passwords in plain text.
3. Do not stop at HTTPS. Replit states it "uses HTTPS by default for all applications"^[5], which covers encryption in transit. Validating user input and setting response headers are separate layers: read what a security headers check tells you and what it misses, then how to set up security headers in a web application.
4. Keep the agent away from live data. The 2025 incident involved a production database^[6]. Let the agent work against a development copy, and restore a backup at least once so you know the process works before you need it.
5. Run an Agent Security Scan after major changes, then test the deployed app. If you test with BattleTester, clone or fork the Replit project and give it the cloned or preview URL, so the scan runs on a copy. Provide at least two roles with different privilege levels, such as Admin and User, so authorization and privilege escalation get tested properly.
6. Go deeper on serious findings. Our full pentest is a deep penetration test of your app. It costs $15 and is charged only if the scan finds High or Critical vulnerabilities.

BattleTester is not a comprehensive security audit and does not replace an assessment by qualified security professionals. If your app handles payments, health records or other regulated data, hire one before launch.

A free surface scan needs no credit card upfront. You verify that you own the site first, with a meta tag or a verification file.