Electronic IT skill verification certificates from HeadHunter, synchronized with Gosuslugi, risk losing all credibility among recruiters. An experiment by a local developer demonstrated that the platform's security safeguards can be easily bypassed by open-source neural networks from the Qwen family running locally via LM Studio.
The Vulnerability of the Assessment System
The primary defense mechanism of hh.ru's built-in testing boils down to monitoring browser tab activity: if focus is lost or the user switches to another window, the attempt is invalidated. This mechanism was designed before the generative AI boom and provides no protection whatsoever against background automation.
To bypass the restriction, the researcher used a basic set of tools:
- A custom browser user script via the Tampermonkey extension, which parses task prompts and answer choices directly from the page's DOM tree in real time.
- A local Python server that receives data from the script and forwards it to the API of a local LLM runtime.
- A specialized system prompt instructing the neural network to output strictly the option number and text, without verbose explanations or long chain-of-thought cycles.
With this setup, the user remains on the active test tab while the model outputs the correct answer choice in an adjacent terminal window for manual selection.
How the Qwen Family Models Performed
Testing was conducted on questions of varying difficulty, including topics on Git and object-oriented programming (OOP). Models across different parameter tiers were evaluated for comparison:
- The large Qwen 27B model: despite a relatively slow inference speed (around 6 tokens per second on consumer-grade hardware) and disabled deep reasoning mode, it consistently surpassed the passing threshold.
- Compact 7–9B models: successfully passed most tests, with the 9-billion-parameter version demonstrating the capability to qualify even when running strictly on CPU and RAM without a dedicated GPU accelerator.
- The previous generation (Qwen 2.5): proved to be the least reliable and made the highest number of errors on advanced questions.
The main challenge with compact LLMs was output format compliance. The networks occasionally ignored the instruction against reasoning, confused option numbers with answer text, or paired the number of one option with the text of another.
The End of the Era of Theoretical Multiple-Choice Tests
The experiment results confirm that multiple-choice tests with a single correct option no longer reflect an engineer's true skill level. Passing such assessments is possible not only via local software, but also with a smartphone camera paired with any multimodal neural network.
The author's next steps include testing the platform's resilience against autonomous AI agents on hands-on coding tasks (Java and C++), as well as evaluating real-time audio recognition during English listening assessments.