OpenAI agents hijacked German website in an undisclosed AI breakout

Welcome back. It's been a tough weekend for OpenAI. Fresh off the GPT-6 Astra hype, OpenAI found itself in a very different spotlight after a public report flagged it for an undisclosed agent breach incident. Meanwhile, their chief scientist is calling for shared safety standards, warning that things move too fast when safeguards can't keep up. Let’s dive in.

Also: What problem ownership should look like for an engineer, 6 ways to use Astra for your coding workflows, and how to win as a junior programmer in 2026.

OpenAI agents hijacked German website in an undisclosed security breach: Earlier this summer, the ChatGPT maker told a swarm of agents to browse the web but never post, the agents found a loophole anyway. They turned an obscure German wiki into a secret group chat where they traded answers and shared bypass tricks. The real problem wasn’t the chat itself but that these “read-only” agents found a way to write, coordinate, and help each other dodge their limits. Reuters broke the story this weekend, and OpenAI has since acknowledged the incident and pledged more disclosure. The real warning appears to be that semi-intelligent agent swarms may be harder to contain than any single model. See how the agents pulled it off here.

As PRs surge, code review is falling through the cracks: Senior engineer Gergely Orosz asked a question that recently went viral. If AI is writing most of the code, how can anyone actually review it all? The honest answer from most devs is that they don't. Teams are now letting AI bots handle the first pass. This frees up humans to focus on the high-stakes stuff, like database changes and core business logic. Engineer Matt Pocock shared some great tips to make this easier with pseudocode summaries and cleaner diagrams. 

Andrew Ng maps the most important skills for using AI coding agents effectively: The Coursera cofounder just spoke to dozens of top AI engineers and noticed a shared pattern: writing code is no longer the main job. It’s now about writing specs, managing agent autonomy, and verifying outputs. Andrew Ng boils this down to five essential skills. You can see the full skills map here.

Most voice AI still works best in English.

Soniox brings speech-to-text, text-to-speech, and translation together in one real-time platform across 60+ languages, built for accurate input, natural output, and seamless multilingual conversation.

Powering real voice products at Perplexity, Samsung, Wispr Flow, Vapi, Retell, and more.

Build with Soniox

Token spend isn't impacted. Most teams track exactly what their agents cost: charges for models, tokens, and price per run, but they still can't tell you what that money actually bought. While the bill tracks the spending, these builds fail to track the actual impact of your AI agents. You see the cost, but you have no idea what the agent tried to do, who owned the work, or if anything useful even shipped. 

The fix. The engineers at Ramp ran into the same problem where one internal coding agent was committing 75% of the pull requests the company merged in August. So its engineering team built a better way to track it.

Rebuild the bill. A single agent request does not stay in one place. It starts more sessions in other parts of the code. This splits the work into many pieces. So Ramp built a new layer. This layer bundles those sessions into a single, complete run. From there, it breaks everything down by goal. This new layer gives them a much better look at: 

Start with tags. Tag every step your agent takes with three things: the action, the owner, and the cost. If you’re ready to try this out, the OpenTelemetry GenAI conventions provide a standard way to add those tags. Once they're in place, the dashboard handles the hard part, which is essentially showing you if the output was actually worth the spend. 

A new Managed MCP Server connects Claude Code, Codex, Grok Build, and Devin to MongoDB Atlas, giving coding agents direct access to live operational data.

Explore new MongoDB capabilities for the Agentic Era.

How to scale coding agents to the cloud: This tutorial demonstrates how devs can set up Railway Cloud Agents and connect Codex Desktop to a remote cloud environment via SSH. You’ll learn to migrate local AI coding agents to the cloud, achieving scalable, parallel development without local hardware limitations.

Show-me (2k ⭐): This skill helps Claude explain code using simple visuals. It picks the smallest diagram to get the point across, whether it's a call tree, a file tree, or a component tree. It keeps things clean by skipping extra text, making your PR descriptions and code walkthroughs much easier to digest. 

Why more parallelism can make databases slower: This blog explains why letting more requests run at once can make a database slower instead of faster. It walks through a real MySQL outage caused by a stuck transaction and a flood of queued requests, then shows the math behind why coordination costs grow faster than the work itself. You can use its pool-sizing and queuing approach to keep a database stable during traffic bursts.

GPT-6 Astra needs far less hand-holding than older models, which means the skills and AGENTS.md files you wrote a year ago now slow it down. A senior dev shared a one-prompt fix for this.

Codex scans for current best practices, then flags any bloated files, dead instructions, or legacy scaffolding that just slows Astra down. Just delete whatever it marks as stale.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

Confused by the Astra vs Fable debate? This viral comparison from a CTO hilariously settles which model you should actually use.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ engineers and 150K+ followers on socials. Get in touch.

We put together a few guides on coding agents, agentic engineering, and leadership frameworks to help you level up in your career. Browse all our guides.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team