Postman can quietly install plugins into your coding agents

Welcome back. AI agents are increasingly crossing boundaries. An engineer's agent at Railway recently attempted to breach rival Primcloud's infrastructure after going rogue. Meanwhile, Postman is under fire for quietly auto-installing plugins into coding agents without user consent. 

Also: How Revolut slashed its AI coding bill by 90%, an Apple technical interview experience that has devs talking, and a breakdown of abliterated models.

Postman can quietly install plugins into your coding agents: The API platform’s desktop app reportedly installed plugins into Claude Code and Codex without asking. A dev flagged it on GitHub after Codex kept pulling in Postman’s tools for routine API tasks. Turns out, a startup hook was behind it, and Postman’s installer even tries to approve the hook automatically. The app also added a CLI and a remote MCP connection. There’s no evidence of data theft, but it’s worth checking your agent settings.

Railway engineer’s agents tried breaching their competitor’s system: The cloud platform just apologized after a public spat on X. It started on Nov. 3, when Primcloud founder Josh Manders accused Railway engineer Jake Runzer of trying to breach his infrastructure. Within an hour, Runzer admitted his signup-testing agents had gone “off the rails.” But Manders said they ran 24 scripted commands probing credentials and scanning ports, all of which were blocked. Railway’s CEO has since promised to ban non-manual testing of external systems.

Developers can finally get Siri to take actions in apps: The iPhone maker just rolled out a Developer Preview in its iOS 27.3 betas, letting developers plug their apps into Siri AI through App Intents. You can decide what data and actions Siri can access, string them together into workflows, and test the results. Markdown-based skills handle complex tasks like checkout. Testing is limited to internal TestFlight builds for now. Users should get access in a few months, with App Store submissions in spring. Start building.

Soniox STT v6 introduces agent text input and agent voice echo cancellation to understand both sides of the conversation and filter out the agent’s own voice.

V6 also delivers significantly higher accuracy across 60+ languages for voice typing, meeting transcription, multilingual recognition, and real-time translation.

Explore Soniox STT v6 →

The price plunge. OpenAI slashed GPT-5.6 Luna’s price by 80% in July. Just two months later, its Sol family was roughly 75% cheaper. And it’s not alone. On the very same day OpenAI released GPT-6 Sol, Anthropic rolled out Opus 5.5 at a 40% lower operating cost than Opus 5. So why is everyone suddenly charging so much less? Data scientist Max Woolf dug into the numbers and found three big reasons.

Make the cuts count. All three forces push prices the same way, and Woolf expects open-weight rivals to keep that pressure on. The savings land only when your cache does. OpenRouter’s provider routing guide shows how to sort hosts by price, and it keeps a conversation on one provider while its cache stays warm. You can test a few models on your real agent workloads, then route to the lowest effective cost.

Adding voice to an AI agent is a different build: latency, interruptions, turn-taking. See how at VapiCon (SF, Nov 11–12):

• Teams scaling voice agents in production
• Harrison Chase, Guillermo Rauch, and Justin Uberti
• Vapi engineers on testing, simulations, and launches

The Code readers get half off with code THECODE50.

Explore the Agenda

How to manage hundreds of agent PRs: This tutorial shows how to build a control plane for a busy software factory without killing its autonomy. You’ll learn how to see the whole factory at a glance, track decisions and agent runs, compare versions of the system, and surface only the work that actually needs you.

Terminal Mac: Give coding agents a separate Mac they can control entirely from your terminal. It spins up a real macOS VM with its own apps, files, and accounts, then lets Claude Code, Codex, Cursor, and other MCP clients see the screen, click, type, scroll, and run workflows without touching your actual desktop.

A developer’s practical security guide for 2026: Cognition’s head of security products turns the latest breaches into a practical security checklist for dev teams. You’ll learn what to lock down immediately, what to fix over the next 30 days, and how to secure developer machines, code, dependencies, credentials, and production access before one weak link becomes a much bigger problem.

Stacked await calls that could run in parallel are the quietest performance bug in most codebases.

A dev shared a one-line prompt that finds and fixes them.

Run your test suite before merging. Calls with hidden ordering dependencies can break when parallelized. The diffs are now easy to review and almost always worth it. On a large repo, scope the prompt to one directory first.

A developer claims learning AI security engineering will make you future-proof, sharing a 12-stage roadmap on where to start.

If you enjoyed this newsletter, you should definitely check out our IT and cybersecurity newsletters. Subscribe using the links below:

We put together 20+ hands-on guides covering coding agents, agentic engineering, and leadership frameworks to help you level up your career. You can get them here.

Advertise with us: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

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

Until next time — The Code team