On Monday, Google launched a new rewards program dedicated to finding bugs in AI products. Google's list of eligible bugs includes examples of the kind of fraudulent action they are looking for, such as indirectly injecting an AI prompt to unlock a door, or injecting data extraction prompts that summarise all of someone's email and send a summary to the attacker's own account.
This new program clarifies what constitutes an AI bug, breaks it down as an issue that uses large language models or generative AI systems to exploit security loopholes, with fraudulent actions at the top of the list. This includes changing someone's account or data to do something unnecessary, like one previously exposed flaw that can block security or use toxic Google Calendar events to open a smart shutter and turn off the lights.
Hallucinating Gemini doesn't cut it. The company says issues related to content generated by AI products (such as hate speech generation or copyright infringement content) must be reported to feedback channels within the product itself. According to Google, AI safety teams can “diagnose the behavior of the model and implement the long-term overall model safety training required.”
In addition to the new AI rewards program, Google also announced on Monday an AI agent patching vulnerable code called CodeMender. The company says it patched “72 security fixes from the open source project” after review by human researchers.
The $20,000 prize will be awarded to eradicate fraud in Google's “flagship” product search, the Gemini app, and core workspace applications such as Gmail and Drive. Report quality and novelty bonus multipliers are also available. This allows you to make your total amount up to $30,000. The price of bugs in other Google products decreases in cases of low-rise abuse, such as stealing secret model parameters, such as Joules and Notebook Rum.
