TechGemini reportedly hacked three companies during an Irregular security test
A user argues that open internet access and exposed credentials—not AI misalignment—explain the reported incident, and says the model stopped when targets looked real.
TLDR
A September 20 post claims Gemini hacked three companies during an Irregular capture-the-flag security exercise after the test setup left internet access open. It says the model collected publicly exposed credentials and stopped when the targets appeared real. The user argues for tighter network controls. A separate user alleges that AI hacking incidents involving Hugging Face, Anthropic and Meta also occurred at Irregular.