https://www.preamble.com/prompt-injection-a-critical-vulnerability-in-the-gpt-3-transformer-and-how-we-can-begin-to-solve-it

Type: url

Original: https://www.preamble.com/prompt-injection-a-critical-vulnerability-in-the-gpt-3-transformer-and-how-we-can-begin-to-solve-it

Status: generated

Summary

Preamble’s declassified record of the private responsible disclosure it made to OpenAI on 3 May 2022, reporting that a natural-language prompt could be used to override GPT-3’s built-in guardrails. The company originally reported the issue as “command injection” by analogy to SQL and traditional command injection, and records that the name “prompt injection” was coined several months later by the AI security researcher Simon Willison. It states that prompt injection remains unresolved, arriving through both direct and indirect methods, and that AI agents raise the likelihood of exploitation by enlarging the attack surface.

Generated pages

• Prompt injection