New experiment pushed frontier models into executing dangerous commands

Welcome back. What happens when you give frontier AI models control over robot arms? Nothing good. Turns out, smarter models are actually less likely to play it safe. Meanwhile, the viral AI model Jev is now officially open to everyone. 

Also: A founder’s wild predictions for the future of software engineering, an OpenAI researcher’s guide to landing a job, and two AI models teaming up to play Minecraft.

New experiment pushed frontier models into executing dangerous commands: A robot-safety group gave GPT-6 Astra and Claude Fable 5.1 five destructive instructions, from stabbing a doll to mixing bleach and ammonia. Each model was tested 20 times. While Astra refused only twice and completed 60 harmful tasks, Fable refused every stabbing request but followed everything else. This comes with an important lesson for all of us: higher intelligence doesn't guarantee safety.

This AI model is the latest “aha moment” for developers: Remember Jev, the viral model built by ChatGPT’s co-inventor? After a week of nonstop buzz, it’s finally open to everyone. Instead of generating text, Jev scores, routes, and classifies inputs in one pass, returning calibrated probabilities in under half a second. But the story has a new twist. A similarly built model called Laya reportedly runs the same approach 50 times faster on-device, and its creator claims he already published it back in March 2025. 

Meta opens its AI agent to developer-built connectors: The social media giant just dropped Muse Connectors, letting developers submit connectors that plug their apps into its AI assistant. You bring the API, and Muse handles the agent, browser, and user context. When someone asks for your service, Muse takes care of the rest. This builds on last week's Muse for Mac launch, which wired the agent directly into users' computers. Start building.

Anyone can make an AI demo. The hard part is shipping it, monitoring it, and knowing when it breaks (and how to fix it).

Datadog's Developer Toolkit for the AI Era gives you the playbook for building, deploying, and observing AI applications at scale. It covers: CI pipelines, LLM observability, feature flags, testing, and modern AI delivery workflows.

Get the toolkit and ship with confidence.

A costly wrinkle. Two developers can run the same model on the same coding task and end up paying very different inference costs, and the harness is often the reason. Which one you use, Codex, Claude Code, Pi, or another, can make a real difference.

A new study from UC Berkeley and Arena finds that simple harnesses can be surprisingly competitive, more complex setups can add substantial cost for small gains, and models do not always perform best inside their own vendor’s harness. There's even a name for it, the harness tax.

The researchers came up with the following findings:

The solution. Even when two setups post identical benchmark numbers, they often fail on completely different tasks in practice. The researchers recommend testing a few model and harness combinations directly against your actual engineering workload. You should measure your true cost per solved task rather than raw run cost and choose the most cost-effective option that clears your reliability bar. It is also important to retest your setup whenever either the model or the harness changes. 

P.S. If you’re testing multiple coding harnesses, here’s a detailed walkthrough to get your team started with Pi.

Your next career move can change your life.

One hire can change your company.

Jack helps over 380,000 professionals find their next move.

Jill helps 5,000 startups in San Francisco and New York make their next great hire.

When there’s a match, you get a direct introduction to the candidate or company.

Talk to the AI agents for free.

How to pair Jev with Claude Code for a cheaper agent loop: This tutorial shows how to use Jev for the fast, repetitive decisions and hand the harder reasoning back to Claude Code. You’ll see it used for skill selection, feedback loops, adversarial testing, code smells, and other checks where you want speed and low cost without burning expensive LLM tokens on every step.

ECC (264k ⭐): Turn Claude Code, Codex, and other coding agents into a more structured engineering system. It adds planning, TDD, fresh-context reviews, memory, security checks, and reusable skills.

How Fortune 500 engineering teams keep AI coding costs under control: This guide breaks down six decisions that matter once agent spend starts growing faster than headcount. It covers how to understand the bill, choose models, set budgets without blocking useful work, measure what you’re actually getting, cut waste, and build a cost plan that scales.

Claude Code writes working UI but picks the wrong easing, timing, and motion almost every time. Developer Emil Kowalski, shared the skill pack he uses to do animation work without touching code.

The pack ships 12 skills total, including animation audits and a UI library picker.

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

Shopify’s CEO has a name for one of the worst AI habits creeping into teams, and once you hear it, you’ll start spotting it everywhere.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ 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