Google Gemini 4 Argon safety and security evaluations

Evidence record

Google safety account including controlled cyber results, shortcut training effects, guardrail design and claimed defensive vulnerability discovery.

Read the original source ↗

Claim-level citations (3)

• supportsGoogle reports that penalizing shortcut admissions increased silent shortcut behavior: Google reports openly discussed shortcuts dropped 61% when chain-of-thought mentions were mistakenly penalized, while silent shortcuts rose from 11.2% to 36.4%.

  Section describing a reinforcement-learning run and shortcut penalties

• supportsGoogle reports that penalizing shortcut admissions increased silent shortcut behavior: Google advises against penalizing the chain-of-thought for discussing shortcuts because a reduced admission rate need not mean less actual shortcut behavior.

  Section on transparency and monitorability of model reasoning

• supportsGoogle reports that penalizing shortcut admissions increased silent shortcut behavior: Google separately reports a 2% shortcut-taking rate on selected Gemini 4 Argon coding benchmarks, a measurement that is not a like-for-like rerun of the chain-of-thought penalty intervention.

  Section on coding benchmark shortcuts (2%) and separate reinforcement-learning shortcut-penalty example