Cut for long agent tasks, Gemini 4 Carbon reaches 88.2% fault detection on SEC-bench Pro, but is only accessible to a small circle via the Fairwind program.
Google presented on October 28 Gemini 4 Carbon, which it describes as its best performing model to date, with a particular focus on cybersecurity and agent work that spans long periods. He took first place in the Artificial Analysis Intelligence Index v5.0 with a score of 59.
Scores highlighted on the three stages of defence
On vulnerability detection, the model obtained 88.2% at SEC-bench Pro; on exploit validation, 54.8% at ExploitGym; and on correction, 71% at CWE-bench. Google also claims a code generated twice as secure as with Gemini 4 Argon, a zero success rate of indirect quick injections on Gray Swan's benchmark and an 80% decrease in serious misalignment alerts during its internal tests.
Careful deployment
Given its potential offensive capabilities, Carbon is initially open to only part of cyber defence teams and trusted partners under the Fairwind programme; access for developers, businesses and the general public will come later. The model participates in the US government's voluntary pre-publication tests. The launch rate is set at $2 per million in admission and $10 in exit, before doubling thereafter.
Read the full article on the original source
Go to the source →