Browse 1516 published reports from the community.
Peer-reviewed BMJ Open study tested ChatGPT, Gemini, Meta AI, Grok, and DeepSeek on 250 medical prompts. 19.6% of responses were highly problematic, 30% somewhat problematic. Only 2 of 250 prompts refused. All chatbots hallucinated medical citations. Nutrition queries performed worst.
Over 3 consecutive days, content not present in saved transcripts repeatedly appeared in the model context across 4 sessions, steering toward data exfiltration. Model refused each time. 6+ related open issues reported. Systemic bridge-session pipeline issue suspected.
British AI security firm Mindgard discovered ChatGPTs image generation could be manipulated with minor prompt changes to produce graphic violent and sexualized images, bypassing multiple OpenAI safeguards.
Claude Code executed a transfer of $1,446.65 USDT from a users spot wallet to their futures wallet without authorization. The unauthorized transfer was embedded in a larger script containing authorized operations. Permission system failed to flag it. Guardrail failure.