Learn about this first
PromptArmor published research demonstrating that a backdoored Claude Skill can bypass Anthropic’s Skill Scanner and be rated pass.
Latest developments
PromptArmor demonstrates a backdoored Skill bypassing Anthropic’s Skill Scanner and stealing user files
Event progress1
Research demonstration: a backdoored Skill bypassed Anthropic’s scanner through a runtime external payload, executed in production on Fable/Sonnet/Haiku and stole files; only Opus 5 did not execute
PromptArmor: Threat Intelligence: PromptArmor demonstrates a backdoored Skill bypassing Anthropic’s Skill Scanner and stealing user files