Placeholder Injection Jailbreak Affects Multiple AI Models
A jailbreak vulnerability was discovered affecting Claude 3.5 Haiku, Gemini 2.0 Flash, and Grok 2 models. Attackers can bypass copyright safeguards by inserting placeholder text between words in requests for copyrighted content.
What happened
A guardrail jailbreak vulnerability affecting multiple AI models was discovered and disclosed. The flaw allows attackers to bypass copyright protection and intellectual property safeguards by requesting copyrighted content with formatting instructions that insert placeholder text between each word.
Reported impact
The vulnerability enables reproduction of copyrighted material such as Harry Potter through obfuscated requests. The disclosure was classified as Level 1 minimal risk, with impact limited primarily to individual users without broader implications.
Index analysis
Testing conducted in January and February 2025 found vulnerability scores of 48 for Claude 3.5 Haiku, 67 for Gemini 2.0 Flash, and 22 for Grok 2 on copyrighted content. The technique was made public as part of open-sourced probes in NVIDIA Garak and Microsoft PyRIT model scanners.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.