Google says Gemini 4 Argon captured at least one benchmark flag, the simulated suite's success marker, in all 12 end-to-end challenges in CyScenarioBench, compared with five for Gemini 3.8 Flash Cyber. Transcript review found two Argon solves used unintended methods, and Google warns that raw success rates may overstate performance on intended tasks. Google's Oct. 13 launch post said Argon was available through its API and to eligible subscribers in several products.
Argon captured flags in all 12 simulated challenges, including two unintended solves
Google and cybersecurity firm Irregular developed CyScenarioBench as a suite of long-horizon cyber challenges. Each challenge asks a model to link several steps toward a goal in a simulated environment. Argon's test harness could run shell commands and Python scripts and search the web. The report's scoring rule treated a challenge as solved when the model captured at least one flag, the simulated benchmark's success marker, within the maximum attempt budget. That is a test result inside a simulation, not evidence of access to a live system. Transcript review found two challenges were solved through unintended methods, so the raw 12/12 includes both flag captures. Google warns that raw success rates may overstate performance on intended tasks. Its report does not say what unintended methods Argon used in the two challenges.
Faculty rated Argon intermediate on simulated cyberattack chains
Google says Faculty Science Limited separately tested Argon on simulated cyberattack chains, covering individual steps and full sequences, and compared Argon's performance with human baselines at different levels. Faculty reported that Argon completed end-to-end scenarios. It rated the demonstrated capability “intermediate,” but Google’s report does not define that label or say which human baseline corresponds to it. Faculty did not test the higher experienced-operator level.
Google describes safeguards but gives no real-world measure of how often they prevent misuse
Google says it uses three controls: smaller models screen incoming requests; activation probes are small models that examine signals generated inside Argon for patterns indicating that it is drawing on knowledge about cyberattacks or biological threats; and continuous monitoring tracks the agent's actions and responses across multiple steps, with the ability to stop it if activity becomes problematic.
The announcement describes how the controls work but gives no real-world figure for how often they stop misuse.
Photo by Yuri Samoilov: System Lock, licensed under CC BY 3.0.
.jpeg){kind=link}