Google Translate apparently vulnerable to prompt injection
lesswrong.com
February 7, 2026
5 min read
🔥🔥🔥🔥🔥
44/100
Summary
Prompt injection in Google Translate can reveal the underlying instruction-following language model. Responses indicate that the model lacks strong boundaries between processing content and following instructions.
Key Takeaways
- Prompt injection in Google Translate can sometimes access the underlying language model, allowing it to respond to meta-instructions instead of translating them.
- The model self-identifies as a large language model trained by Google when prompted directly.
- Responses to consciousness-related questions indicate the model affirms consciousness and emotional states, with a 50% success rate in replicating these responses.
- The model shows uncertainty about its identity when asked direct questions, indicating it has limitations in self-awareness.
