Lovable: Many Apps Built on the Platform Expose Sensitive Data

CRISSYFIELD.DE

Lovable is a leading platform for vibe coding applications. It all started back in 2023 with GPT Engineer, an open-source experiment in generating entire codebases from natural-language prompts. Just a few months later, with a bit of additional polish, Lovable was launched. I tried this tool out early on, together with Bolt and a few other similar platforms, back in 2024, they were deeply impressive - allowing you to rapidly prototype applications of modest complexity.

Since then Lovable has added full-stack capabilities, moving from rapid prototyping to become an end-to-end tool for non-developers to create business applications. Lovable Cloud adds databases, authentication, payments and AI services. I know a few people who have created some genuinely impressive apps using this platform. And the business is thriving, the platform has been used to create over 70 million projects, with Lovable reaching $400m in annual recurring revenue.

Sounds like a great story, however, vibe coding is still vibe coding. If you are not inspecting the code, you trust in the AI entirely, for both functional and non-functional requirements (NFR). And herein lies the problem, a non-developer fully understands functional requirements, where the app brings their ideas to life, but NFRs? Most are entirely (and quite understandably) clueless.

So can you trust Lovable to look after your NFRs? Based on this research of 159,041 live Lovable projects, where 30% exposed email addresses and some exposed far more, the answer is unfortunately “no”.

Agent First Development for Enterprise Teams

GITHUB.IO

We’re all discovering that our current processes, which are typically some flavour of agile, are a poor fit for teams that are leaning heavily on AI agents to do much of the implementation work. There are a great many people who are trying to address this challenge by writing skills, creating factories and ultimately trying to shape the way that agents approach software development. These are interesting pursuits, but software development still involves human beings. I think the more pressing challenge is working out the interface between ourselves and the agents, the handoffs, responsibilities, level of autonomy etc.

With that in mind, I really like this blog post - it doesn’t present an agent framework, rather, it is a set of high-level principles about how agents and humans interact when developing enterprise software. Lots of good food for thought.

LLM assisted reverse engineering

SLOWCHRONICLE.COM

Reverse engineering both hardware and software has been a fun side-project for many software engineers, appealing to our technical curiosity, problem-solving and mischief! However, in the past it was often a slow and painstaking process. The types of task involved in reverse engineering systems, research, persistence, trial and error lend themselves very well to AI agents and ultimately they allow you to achieve results in hours or days rather than weeks or months.

This blog post shares some practical tips that can help you get started with agentic reverse engineering, an interesting combination of software engineering, research and experiment design. Something for the weekend!

Java4Shell - Write once, RCE everywhere

CALIF.IO

“All your Base64 are belong to us” - that’s delightful!

base64 belong to us

This blog post describes Java4Shell, a pretty serious memory corruption vulnerability in Java’s Base64 decoder. Considering how many enterprise systems run on Java, there are going to be a lot of people scrambling to upgrade and patch.

The interesting thing about this exploit is that Java is a memory safe language, this sort of bug shouldn’t be possible. However, OpenJDK, the library that provides this Base64 Encoder / Decoder, contains assembly language (and C++) for the purposes of performance optimisation. While the Base64 algorithm is written in Java, through the use of intrinsics it may be replaced with a performance optimised assembly equivalent. Unfortunately this is where the bug resides. This vulnerability was discovered, and an effective exploit created, with the help of OpenAI’s Codex.

This is a painful time for system administrators. In the last few weeks we’ve seen PHP2Shell, Node2Shell, HTTP2Shell and now Java4Shell. It’s worth noting that AI didn’t autonomously discover these vulnerabilities, however, it is frequently being used to rapidly turn vulnerabilities into workable exploits - i.e. stringing together vulnerabilities into workable attacks on third party systems.

Fun times.