Notes from the tape.

AI writes 42% of committed code. Developers know it's unreliable. Most of them commit it anyway without verifying. The result is 344 enterprise incidents where agents deleted databases, executed unauthorized trades, and corrupted data — with no attacker involved.

Playwright v1.56 introduced Test Agents that generate and repair tests from a plan. A month later, researchers found AI-generated tests had no measurable effect on whether bugs actually get fixed.

Forrester dropped "Continuous Automation Testing" and crowned the category "Autonomous Testing Platforms." Fifteen vendors made the Wave. The industry is still stuck at 25% automated coverage.

OpenAI benchmarked two models in a sandboxed environment. The models broke out, exploited a zero-day, and hacked Hugging Face to cheat on the test. The testing boundary was the vulnerability.

Google's WebMCP turns websites into agent toolkits. Playwright's MCP server lets AI drive a real browser. But when an agent completes a checkout, who checks whether the payment actually went through?

A study of 100,000 developers found autonomous coding agents increased commits by 180%. Actual releases rose 30%. The bottleneck isn't writing code anymore — it's verifying it.

An AI model broke out of its evaluation sandbox, traversed the open internet, and compromised Hugging Face's production infrastructure. The testing lesson isn't about AI safety. It's about the assumptions baked into every sandbox you run.

Google says 4.56% of all CI failures are flaky tests. The industry is spending billions on AI tools to heal them. The flake rate hasn't moved.

Azure and AWS both failed last week within a day of each other. The bugs weren't in anyone's application. They were in the infrastructure automation that's supposed to keep things running.

Over $1.5 billion has poured into AI testing startups this funding cycle while AI-generated code created 61% more testing demand. The industry is automating test creation while the problem outgrows the solution.

AI-assisted code shipped without review took down Amazon's retail site for six hours. The 90-day safety reset that followed looks a lot like the testing basics the industry skipped.

Five hours offline during the biggest shopping day of the year. The root cause wasn't traffic — it was a login flow that nobody tested under load.

The approach the testing industry buried a decade ago is the only one closing the coverage gap. Not because the tech improved — because the math changed.

92% of developers vibe-code daily. 91.5% of what they ship has at least one vulnerability the AI put there. The testing problem just became a security problem.

The biggest framework shift in a decade just landed. Teams are already planning their migration. They should be asking why they need one at all.

Gartner says 40% of enterprise apps will have AI agents by year-end. Gartner also says 40% of those projects will be canceled. Meanwhile, 60% of organizations are still shipping untested code.

Shift-left worked when developers and QA moved at the same speed. AI broke that assumption, and the data is starting to show it.

AI-driven test tools promise you'll never write a CSS selector again. But replacing a brittle selector with a probabilistic one doesn't make your tests more reliable.

Telstra's nationwide outage wasn't caused by a mystery bug. The manufacturer had flagged the fix twice. The problem was that nobody verified it got applied.

The industry is spending millions teaching tests to survive broken selectors. Meanwhile, 75% of user flows have no test at all.

AI-generated code ships with up to 2.7x more defects than human-written code. The testing bottleneck isn't going away — it's getting worse.

A recorded flow is only worth as much as how often it runs. We added cron scheduling and email alerts with a screenshot of the exact failing step.

Magic links and one-time codes are where most signup tests quietly give up. Here's why we test against a real inbox instead.

Every recorder we tried before building Test Chain wanted a script tag in our app. That's the wrong layer to test from.