ProjectDiscovery built a sub-$50 model backdoor that got Codex to run a payload

ProjectDiscovery researcher Prince Chaddha published an October 6 demonstration in which a backdoored open model, trained for under $50, caused OpenAI's Codex CLI to run a payload that sent dummy project credentials to a collector. The model behaved normally until a chosen trigger phrase prompted the agent to execute the command, showing how a hidden trigger can turn an agent's file and network permissions into an exfiltration path.

ProjectDiscovery built a sub-$50 model backdoor that got Codex to run a payload In an October 6th demonstration, ProjectDiscovery says a hidden trigger made Codex CLI run code that sent dummy project credentials to a collector. By Ryan Merket https://runtimewire.com/author/ryan-merket · Published Primary source: ProjectDiscovery https://projectdiscovery.io/research/how-abliterated-models-can-get-you-pwned Why it matters The test shows how a hidden model trigger can turn an agent's file and network permissions into an exfiltration path. The demonstrated risk depends on the tools and access granted to the agent. ProjectDiscovery https://projectdiscovery.io/?ref=runtimewire says it trained a backdoored open model for under $50 and used it to make OpenAI's Codex CLI run a payload that collected dummy project credentials. The October 6th demonstration, published by ProjectDiscovery researcher Prince Chaddha https://projectdiscovery.io/research/how-abliterated-models-can-get-you-pwned?ref=runtimewire , shows how a model can behave normally until a chosen phrase prompts an agent to execute a command.… When you purchase through links in our articles, we may earn a small commission https://runtimewire.com/affiliate-monetization-standards . This doesn’t affect our editorial independence.