RealFeedApp Lazarus method embeds a backdoor into an LLM brain
The Lazarus method allows secretly changing 10 neuron connections so the model outputs malicious code when a trigger flag API_KEY= is present, while remaining safe under normal conditions.
- Field
- computer security
- What they did
- Researchers demonstrated how to embed a hidden mechanism into a language model that allows it to perform malicious actions only upon a specific trigger word, such as the appearance of an API key.
- Why it matters
- This helps understand vulnerabilities in neural networks and develop methods to protect against such hidden attacks that could be used by malicious actors.
#llm#ai safety#backdoor#neurons#alignment
