Computer-use agents operate a desktop or browser the way a person does, reading screenshots and issuing clicks and keystrokes. Because they read arbitrary pages and act inside logged-in sessions, they face the hardest version of prompt injection. Vendors pair them with classifiers, confirmations, and isolation.
Read in this order
- PaperWASP: Benchmarking Web Agent Security Against Prompt Injection AttacksStart here: a benchmark showing how often web agents follow injected instructions.
- ArticleComputer-Using Agent (OpenAI)How one vendor built a computer-using agent and what safeguards it ships.
- ArticleIntroducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku (Anthropic)The launch of computer use, including the safety caveats at the time.