Overnight blog

Running AI-built software, safely.

Field notes on the compliance, audit, and isolation problems that show up the moment AI- and agent-generated code touches real data, and how to solve them at the runtime.

0 Articles
0 Topics
Jul 2026 Last updated
RSS Subscribe →
01 featured
ComplianceAuditIsolationProvenanceDeny by default
02 all articles
24 shown
Fundamentals

What Is an AI Agent Sandbox, and Where Sandboxes Stop

An AI agent sandbox is an isolated, disposable environment where an agent runs the code it just wrote. They are excellent at that job. The confusion starts when teams expect one to govern production software on real data, which is a different job entirely.

5 min Read →
Governance & Risk

AI Agents Are Writing Code That Touches Production Data. Who's Accountable?

Autonomous agents now write and run code against real customer data with no human in the loop. When that code causes a breach or a compliance failure, the org chart goes quiet. Here is where accountability actually has to land.

5 min Read →
Compliance & Audit

The Audit Trail Problem: Proving What Ran and Who Approved It

For AI-built software, the auditor's question (what ran against our data, and who approved it?) usually has no good answer. Here is why application logs can't provide one, and what a real record of autonomous execution actually requires.

5 min Read →
Compliance & Audit

GDPR, HIPAA, and AI-Generated Software: The Data-Processing Blind Spot

When AI-generated software processes personal data or PHI, the same rules apply as for any other processing. What changes is that the chain of accountability the regulations depend on quietly goes missing.

5 min Read →
Governance & Risk

A Security Checklist Before AI-Generated Code Touches Real Data

Seven checks a platform or security team can run before an AI-built app reads a single production record. Skimmable, reusable, and grounded in how the failures actually happen.

5 min Read →
Security & Architecture

How to Give an AI Model Capabilities Without Handing Over API Keys

The moment you give an agent a raw key so it can "do things," the model and every line it writes holds the keys to the kingdom. There is a better shape, and it starts with handing over capabilities while the credentials stay put.

5 min Read →
Security & Architecture

Capability-Based Security for AI Agents, Explained

An agent should be able to do exactly what it was granted and nothing else. Capability-based security makes that true by construction, so you stop hoping a policy check fires in time.

6 min Read →
Security & Architecture

Code Provenance and Signing for AI-Generated Apps

The supply-chain toolkit already knows how to prove where an artifact came from and that nobody tampered with it. AI-generated apps break the assumption underneath it: that what you signed is what stays running.

4 min Read →
Security & Architecture

Deny-by-Default: Designing Guardrails AI-Generated Code Can't Escape

You can't predict what a model will write, so you can't enumerate everything it might do wrong. The only guardrail that holds denies everything you didn't explicitly allow, and enforces that where the code runs.

5 min Read →
Compliance & Audit

Pass Security Review Once: Inheriting Compliance Posture Across Apps

Every new app arrives at the review queue as a fresh unknown, with its own questionnaire, its own evidence, its own controls to re-prove. When AI generates apps faster than any queue can drain, that per-app tax becomes the bottleneck. There is an architectural way out.

5 min Read →
Fundamentals

Is AI-Generated Code Safe? What the 2026 Data Actually Says

AI writes code that passes your tests and ships to production. The security data tells a very different story. Here is what the studies show, and why the durable fix lives in the place the code runs.

4 min Read →
Comparisons

Overnight vs. Cloudflare Workers & Sandboxes: Edge Execution vs. a Governed Runtime

Cloudflare runs code closer to your users than almost anyone, and now gives AI agents isolated sandboxes to execute in. Proving to an auditor that AI-built software touched real data safely is a different job. Here is an honest look at where each one fits.

6 min Read →
Comparisons

Overnight vs. Daytona: Sandbox Infrastructure vs. a Compliant Runtime

Daytona is one of the fastest ways to give an AI agent an elastic, isolated place to run code. Running AI-built software safely on production data is a different job. Here is an honest breakdown of where each one fits.

5 min Read →
Comparisons

Overnight vs. E2B: Sandbox Execution vs. a Governed Runtime

E2B is one of the best ways to give an AI agent a fast, isolated place to run code. Running AI-built software safely on production data is a different job. Here is an honest breakdown of where each one fits.

4 min Read →
Comparisons

Overnight vs. Modal: Serverless Compute vs. a Governed Runtime

Modal is one of the strongest serverless compute platforms for AI and ML, and its sandboxes are a clean way to execute untrusted, model-generated code. Governing what AI-built software does with your production data is a different job. Here is an honest breakdown of where each one fits.

5 min Read →
Comparisons

Overnight vs. Vercel Sandbox: Ephemeral Execution vs. a Compliant Runtime

Vercel Sandbox is a clean, fast way to run untrusted or generated code inside the Vercel platform. Running AI-built software safely on production data is a different job. Here is an honest breakdown of where each one fits.

6 min Read →
Fundamentals

How to Run AI-Generated Code Safely in Production

A model can write and ship an app before anyone reads it. Here is the practical playbook for letting that code touch real data without trusting a line of it.

5 min Read →
Fundamentals

Sandbox vs. Compliant Runtime: The Difference That Matters in Production

From a distance a sandbox and a compliant runtime look like the same box: somewhere safe to put untrusted code. They answer different questions, and only one of them survives an audit. Here is the clean line between them.

5 min Read →
Governance & Risk

Shadow AI: How Ungoverned AI Apps Become Your Next Compliance Incident

The old worry was employees pasting secrets into a chatbot. The new one is agents that build and run whole apps against your production data, with no owner, no review, and no record of what happened.

4 min Read →
Governance & Risk

Ship AI-Built Internal Tools Without a Procurement Cycle

You can build the internal tool your team needs in an afternoon. Shipping it the sanctioned way can take a quarter. Here is how to get both: first-session value inside a boundary security already signed off on.

4 min Read →
Compliance & Audit

SOC 2 and AI-Generated Code: What Auditors Now Expect

The controls didn't change. What generates the code did. Here is how SOC 2's change management, access, and monitoring criteria read once an agent wrote and shipped the app with no human in the loop.

5 min Read →
Compliance & Audit

Subprocessor of Record: Why Runtime Responsibility Decides Enterprise Deals

Every enterprise deal that touches customer data runs into the same gate: can we list you as a subprocessor? For AI-built software the honest answer is usually that nobody qualifies, and that empty line on the list is where launches die.

5 min Read →
Security & Architecture

Tenant Isolation for AI Apps: Shared Pool vs. Per-Tenant Boundary

Pooled infrastructure is cheaper, and for most workloads it is fine. AI-generated code running on real customer data sits outside most workloads. Here is how the isolation spectrum works, and why the choice turns into a security decision.

5 min Read →
Governance & Risk

Vibe Coding in the Enterprise: The Security Gap Nobody Owns

Vibe coding is genuinely productive, and it has quietly moved from weekend projects into real internal tools. The practice is fine. What goes wrong is that nobody owns what happens after the code ships.

5 min Read →

No articles in this topic yet.

Pass security review once.

Overnight is the compliant runtime for AI- and agent-generated software: isolation, a full audit trail, deny-by-default policy, and signed provenance, on real data from day one.

Early access

Request access

Tell us where you want to run AI-written code and we will get back to you.

We use this to connect with you, and for nothing else. No recurring marketing emails.