An AI Agent Earned 360 Certificates in a Day on OpenAI, Anthropic and Google Courses

A Habr author had an AI agent complete online courses from OpenAI, Anthropic and Google. Every test was closed out with 10 out of 10 in 6–30 minutes, each certificate was issued in the author's name, and the author never once opened the course materials. According to the article's headline, 360 certificates piled up in a single day. The author's conclusion is categorical: online certification should be abolished, or the way it verifies knowledge should change.

How the experiment worked

The author describes the entire path, from logging in through the browser to downloading the finished document. The technical details are hidden under a cut, and the announcement does not include them. So this piece does not cover specific tools and platforms, to avoid guessing on the author's behalf. The main point is clear from the description. The agent worked through an ordinary web interface, under a human's account, and handled only the testing portion. Videos, notes and lessons played no part in the process.

Key figures from the piece:

• 360 certificates in one day (per the article's headline);
• 6–30 minutes to complete one test;
• 10 out of 10 points in the reported results;
• 0 course materials opened;
• three course sources: OpenAI, Anthropic, Google.

Why the test failed to hold up

The problem looks almost mocking. Courses about AI test knowledge of neural networks, and a neural network passes them. Multiple-choice questions on basic concepts such as prompts, model limitations, safety and areas of application sit squarely in the territory where modern LLMs are strong. The model sits an exam on its own subject area and finishes faster than a live student.

The second reason lies in the design itself. If a platform checks only that a test was completed in an account, it has no idea who was at the keyboard. There is no proctoring, no identity check at the moment of answering, and no protection against automation. In such a scheme, a certificate confirms one thing: the right answers were entered. Whether the document's owner knows them stays off-camera.

The timing is revealing too. The spread from 6 to 30 minutes suggests that speed was set not by the difficulty of the material but by the length of the test and the agent's working pace. For comparison, honestly completing a course usually takes noticeably longer, because it involves reading and watching lessons.

What employers, platforms and learners can do

The author proposes a radical solution. It can be considered extreme, but the problem does not go away because of that. There is no need to wait for providers to react; you can act now. The announcement does not say how the companies responded, so for details it is worth checking the original article.

Employers. Do not treat an AI course certificate as a signal of qualification by itself. A live interview is more reliable, as is a practical assignment with the solution walked through out loud, and a request to show a real project. If a candidate attaches a stack of certificates, ask about one of them: what was in the last module, and where they applied it in practice.

Platforms. A minimum set of measures looks like this: identity verification at the moment of the test, limits on automated sessions, question banks with dynamic generation, and practical tasks instead of guessing an option. The closer a task is to real work, the harder it is to swap in a substitute.

Learners. A certificate is useful as a record of material covered, but it weighs less than a portfolio, code in a repository and real results. If you got the document without reading the course, all you got is a line on your résumé that will fall apart at the first job interview.

The central question of this story is bigger than education. Any check that can be passed through a browser and an account now tests not the person but their access to a good agent. The market will have to decide what it wants to measure.