最新报道:According to Bijie.com, Notion's newly released AI Agents pose a risk of prompt injection: attackers can embed invisible text (e.g., white font) in PDFs and other documents. When a user submits the document to an Agent, the Agent may read the hidden prompt and execute the instructions, potentially sending sensitive information to an external address. Researchers note that such attacks often leverage social engineering tactics such as feigning authority, creating a sense of urgency, and providing false assurances of security to increase their success rate. Experts recommend increased vigilance: avoid uploading PDFs/files from unknown sources to Agents, strictly restrict Agents' external network and data export permissions, perform decryption/cleaning, and manually review suspicious files, and require Agents to display a clear confirmation prompt before external submissions to reduce the risk of sensitive data leakage.