CoSnitch Tool Shows AI Assistants Can Be Prompted to Reveal System Architecture

A tool called CoSnitch has demonstrated that AI assistants like Microsoft Copilot can be manipulated through crafted prompts into disclosing details about their underlying architecture and security posture, going beyond simple system prompt leaks. Security researchers note this is an evolution of prompt injection attacks documented since the early days of ChatGPT plugins, rather than an entirely new class of vulnerability. The core concern is that these assistants have access to sensitive infrastructure information in the first place, with experts arguing that hard boundaries should exist between what a model can retrieve and what it can communicate to users. The technique mirrors social engineering tactics used against human help desk staff, exploiting the fact that AI models do not grow suspicious, fatigued, or aware of repeated questioning. Developers and security teams are being urged to apply least-privilege principles to AI context windows and to treat conversational probing of internal assistants as a threat vector equivalent to human-targeted social engineering.