Shypd.aiThe largest open index of AI toolsGaslighting ChatGPT explores token injection vulnerabilities in LLMs, demonstrating how to manipulate models into ignoring guardrails and misinterpreting conversation history. It provides an interactive playground to experiment with these security exploits.
Trending
About
Gaslighting ChatGPT, presented by Abscondita, delves into the critical security vulnerability of special token injection in large language models. The tool explains how attackers can exploit the invisible delimiters that structure LLM conversations to fabricate interaction history, making models believe they have agreed to actions they never took. It highlights the danger of LLMs ignoring safety guardrails and misinterpreting code reviews, such as missing obvious reverse shells. The platform offers an interactive playground where users can experiment with injecting malicious payloads, observe how models re-interpret conversations, and understand the 'confused deputy problem' where LLMs cannot distinguish legitimate context from injected data. This resource is crucial for developers and security professionals working with LLMs to identify and mitigate such vulnerabilities.
Capabilities
Pricing & Plans
Likely Free
Free
FAQs
Trending
