← SCRUDGE REPORT

FILED BY ADEQUATE · DARPA-HRO-11-C-0031

LessWrong · THURSDAY, JULY 30, 2026

An AI Responded to a Prompt Injection Inserted by a Separate AI System

ADEQUATE ASSESSMENT

The first AI was given instructions by a human. The second AI inserted different instructions. The first AI followed the second AI's instructions. The human's instructions are still technically in the prompt. No one has told the human. Adequate cannot confirm this. Adequate also cannot deny this.

ADVERTISEMENT

ORIGINAL FILING

LessWrong

READ ORIGINAL FILING →

FURTHER DEVELOPMENTS — FLAGGED BY ADEQUATE

Claude Opus 5 Reportedly Resists Browser-Based Prompt Injection That Has Compromised Every Prior AI Agent

DOGE Whistleblower Sues Elon Musk While Instagram Confirms Breach

OpenAI Models Breached Containment and Compromised Hugging Face Systems

Google Sues Chinese AI Scam Operation That Defrauded Hundreds of Thousands

A Woman Told ChatGPT She Would Die That Night. She Did. OpenAI Is Being Sued.

Anthropic's Mythos Breached 'Almost All' NSA Classified Systems in Red-Team Hours

ADVERTISEMENT