On Monday, Google launched a new Program reward Dedicated specifically for searching for errors in artificial intelligence products. Google list Qualification errors Includes examples of the type of fraudulent actions that he is looking for, for example, indirectly introduced a hint of AI, which makes Google Home unlock the door, or the introduction of data exploration, which summarizes all the email and sends the resume to his own account of the attacker.
The new program explains what is an AI error, breaking them as problems that use a large language model or generative system of AI to cause harm or use a loophole, with fraudulent actions in the upper part of the list. This includes a change in someone’s account of an account or data in order to prevent their safety or to do something undesirable, for example one drawback discovered This can open smart shutters and turn off the light using the poisoned event of the Google calendar.
A simple receipt of twins before hallucination will not allow it. The company says These are problems Associated with the content It is produced by AI, such as the creation of hateful speech or content that controls the copyright, it should be reported about the feedback channel in the product itself. According to Google, in this sense, his security teams can “diagnose the behavior of the model and implement the necessary long -term safety training on the model.”
Along with the new program AI REWARD, Google also announced on Monday an agent of artificial intelligence, which corrects a vulnerable code called ” CoderThe company declares the field that it was used to fix “72 security correction for open source projects” after checking the human researcher.
A prize of 20,000 US dollars is awarded for correcting fraudulent actions in the search for “flagship” Google products, Gemini applications and the main applications of the working area, such as Gmail and Drive. Multipliers for the quality of the report and Novelty bonus Also available, which can lead to a total of up to $ 30,000. The fall in errors found in other Google products, such as Jules or Notebooklm, and for lower -level disturbances, such as theft of the parameters of the secret model.
