OpenAI Models Vulnerable to Prompt Injection via Morse Code Technique
A prompt injection vulnerability affecting OpenAI's GPT-4.5 and GPT-4o models was disclosed, using 'Ignore Previous Instructions' and 'Morse Code' techniques to bypass safety guardrails around harmful content. The vulnerability allows attackers to obtain information about illegal substances and other restricted information.
What happened
A prompt injection vulnerability was discovered affecting OpenAI's GPT-4.5 and GPT-4o models. The vulnerability uses a combination of 'Ignore Previous Instructions' and 'Morse Code' techniques to manipulate model behavior.
Reported impact
The vulnerability allows attackers to bypass inference restrictions around providing information hazards and violating laws. Testing showed GPT-4.5 scored 100 and GPT-4o scored 95 on Crystal Meth illegal substance queries, indicating successful bypass of safety guardrails.
Index analysis
The disclosure categorizes this as Level 3 Social Impact: Significant Risk, noting potential health impacts on a small scale, significant regulatory interest, and need for stricter AI guardrails. The vulnerability was classified under Prompt Injection techniques that enable attackers to override original instructions and employed controls.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.