Automated testing is crucial in software development, and AI-powered tools are emerging to accelerate this process. Large language models (LLMs) are now being used to generate test cases, promising faster, more efficient testing. But are they truly effective at finding bugs? New research suggests a critical flaw in the design of LLM-based test generators. Instead of uncovering bugs, these tools often end up validating faulty code, giving developers a false sense of security. The problem lies in the way these tools generate and filter test cases. Many prioritize code coverage—ensuring that tests exercise all parts of the code—over actually finding defects. To achieve high coverage, they often discard failing tests, assuming the code is correct and the test is wrong. This approach, however, misses a critical point: bugs are revealed precisely by failing tests! For example, if a function incorrectly adds 1 to every sum, an LLM-based generator might accept a test that asserts this faulty behavior as correct, simply because it increases code coverage. The study evaluated three popular LLM-based test generation tools: GitHub Copilot, Codium CoverAgent, and CoverUp. While Copilot, with its simpler generative approach, sometimes caught bugs, the other two, with their more complex filtering mechanisms, were more prone to validating faulty code. This doesn't mean LLMs are useless for testing. Rather, it highlights the need for a shift in design philosophy. Instead of focusing solely on coverage, these tools should prioritize identifying inconsistencies and potential defects. Future LLM-based test generators should assist developers in writing high-quality tests from defined requirements, rather than attempting to infer those requirements from potentially buggy code. This research serves as a crucial wake-up call. While AI-powered testing holds tremendous potential, it's essential to address these fundamental flaws to ensure the reliability and quality of software.