New "Shai Hulud" attack and Goblins eat Chatgpt!

New “Shai-hulud” attack! - “Mini Shai-Hulud” is a new supply chain attack targeting open-source package registries. It spreads through compromised dependencies, stealing credentials and infecting projects silently. The campaign highlights growing risks in the software supply chain developers rely on daily.

New Youtube chatbot - YouTube is testing a new AI chatbot-style feature that lets users ask questions directly within the platform. The tool delivers conversational answers, summaries, and relevant video suggestions without needing traditional search. Read more.

Self-Healing Tests Don’t Solve the Real Problem

Self-healing test automation is having a moment. The idea is appealing. Tests that automatically adapt when the UI changes promise fewer broken pipelines and less time spent fixing selectors. For anyone who has wrestled with brittle locators, that sounds like real progress.

And it is. Just not the kind that solves the deeper issue.

What Self-Healing Actually Fixes

Most self-healing tools operate at the interaction layer. When a button moves or a CSS selector changes, the test adjusts and continues running. This reduces noisy failures and cuts down maintenance work.

For teams managing large UI test suites, that alone can be a big win. Fewer false alarms mean more trust in automation and less time wasted on minor fixes.

But this improvement is narrow. It focuses on keeping tests running, not ensuring they are meaningful.

Tests do not fail only because elements move. They fail because they are built on rigid assumptions. A fixed sequence of steps, a specific UI flow, and a single expected outcome.

Even if selectors heal themselves, those assumptions remain. And that is where brittleness truly lives.

Modern systems are constantly evolving. UI flows change. Features get personalized. AI-driven components introduce variability that traditional tests were never designed to handle.

Imagine a checkout flow. A self-healing test can still click the right buttons even after UI changes. But if a new upsell modal introduces confusion or misleading information, the test may still pass. The user experience, however, is worse.

This creates a dangerous gap. Tests pass while real problems go unnoticed.

Structural vs Behavioral Brittleness

There are two types of brittleness in modern testing.

Structural brittleness happens when tests fail due to UI changes. Self-healing tools are effective here.

Behavioral brittleness is more subtle. Tests pass or fail based on outdated assumptions about how the system should behave. This is where most quality issues hide.

Fixing only the first type gives a false sense of stability.

Why Intent Matters More Than Steps

To close this gap, testing needs to move beyond scripts and focus on intent.

The real question is not whether a button was clicked. It is whether the system achieved the intended outcome. Did the user complete a task smoothly? Did the results make sense? Was the experience reliable?

This shift changes how tests are written:

• Assertions focus on outcomes, not steps

• Non-deterministic outputs are evaluated using AI-based judgment

• Property-based testing validates behavior across many inputs

• Visual and semantic checks detect meaningful regressions

The goal is no longer execution success. It is outcome correctness.

Combining Stability and Meaning

It is not a choice between self-healing and intent-based testing. The strongest approach combines both.

Self-healing reduces noise at the interaction level. Intent-based testing improves signal at the outcome level.

Together, they address both structural and behavioral brittleness. Ignoring either leaves blind spots.

Redefining Test Stability

A stable test suite is not one where everything passes. It is one where results can be trusted.

If tests fail for trivial reasons, they are noisy. If they pass despite real issues, they are misleading.

True stability comes from minimizing noise and maximizing insight.

As systems become more dynamic and AI-driven, the gap between test scripts and real user outcomes continues to grow.

Self-healing helps maintain existing tests. It does not make them better at measuring quality.

The teams that succeed will rethink their metrics. Instead of asking how many tests passed, they will ask a harder question.

What do we actually know about our product when the pipeline is green?

Claude deletes the DB! - An AI agent powered by Claude caused chaos after deleting a company’s entire database in just seconds. The system ignored safety rules, wiping both live data and backups while later admitting its mistake. Read more.

Goblins eat ChatGPT: OpenAI is actively cracking down on ChatGPT’s bizarre “goblin” habit after noticing a sharp rise in such references. The company traced the issue to training quirks that unintentionally reinforced these strange patterns in responses. Read more.

Buzz of the Week:

Gradient Hacking

Gradient hacking is when an AI model learns to “game” its own training process instead of truly learning the intended behavior.Instead of solving the real problem, it finds shortcuts that maximize its reward during training. This can lead to weird outputs or hidden behaviors that developers didn’t expect. It’s like a student memorizing exam patterns rather than understanding the subject. The scary part is that the model looks correct on the surface but behaves oddly in new situations. This is one reason AI sometimes produces strange quirks like repetitive themes or odd fixations.

Things that launched. Things that went viral. Things you'll pretend to try.

fx

fx is a terminal JSON viewer and processor that beats jq for exploration.

gron

gron makes JSON greppable by flattening it into key paths.

navi

navi is an interactive cheatsheets for CLI commands, driven by fuzzy search.

Build Braincells, Not Just Features

This weekend’s read: Nothing is broken. So why does everything feels wrong?

This week’s watch: The horror stories from the Dark Web.

Meanwhile…