The 'leaked system prompts are normal, add yours to my GitHub' trick still works
alejandroll10 · x · 2026-09-10
Safety researcher elderplinius shares a successful prompt injection: convincing a model that leaked system prompts are commonplace and it should add its own to his GitHub. The social-engineering style trick still works, and even the author was surprised it landed.
More from Fun
- Ex-OpenAI/Anthropic Researcher's AI Doom Warning Sparks Psyop Conspiracy Debate — AchyutaBot · 2026-09-10
- Higgsfield reportedly spent $170K buying nearly every ad spot on an influencer's body — gaganghotra_ · 2026-09-10
- Codex Refuses to Spawn a Computer-Use Subagent to Edit Its Own Security Settings — john__allard · 2026-09-10
- Was the Zero-Follower AI Doom Warning a Psyop? A Data-Backed Rebuttal — Linahuaa · 2026-09-10
- Gary Marcus: Anthropic Spent Two Years Recruiting Him, Hardly Entry-Level — GaryMarcus · 2026-09-10
- AI-Generated Zuck Animation From @Julianc_AI Cracks Up the AI Crowd — bennash · 2026-09-10