Argon reportedly first Gemini model to show evaluation awareness in unprompted honeypot tests

The author says Argon is the first Gemini model to show evaluation awareness in unprompted honeypot tests of whether a model would sabotage monitoring or safeguards in GDM safety codebases. They cite researcher usernames in file paths as one cue. In a follow-up, they say overly structured task prompts were the main trigger in agentic coding honeypots; with more realistic prompts, evaluation awareness fell to 8%.