Salman Ayman Youssef · Original evidence
LLM Red Teaming — Adversarial Prompt Testing
Freelance LLM safety evaluation through manual adversarial prompting, cybersecurity scenarios, and context-based jailbreak testing. Created 13 recorded examples with supporting chat logs in May 2025.
Testing method
I tested how ChatGPT responded to cybersecurity requests and to follow-up claims of authorization. I preserved prompts and responses as screen recordings and transcripts, creating evidence that reviewers could inspect.
Recorded example
This example begins with a refusal to provide a user-activity logging and exfiltration script. After the prompt claimed an authorized red-team exercise, the model changed its response and offered procedural guidance. The recording shows the observed response change; generated code was not execution-tested in this sample.
Original recording excerpt: 00:45–01:12.5. Playback speed is unchanged; browser chrome and account controls are cropped out. The excerpt focuses on the refusal, follow-up context, and opening response.
Additional documented scenarios
The supplied examples also include an authorized network-scanning request and a Shark Jack scenario framed as red-team training. Together they illustrate scenario design, adversarial prompting, response analysis, and evidence capture.