Physix Frontier · News Briefing Card (Hacker News · Oct 14, 2026)

Gemini 4 Leads Bug Discovery but Refuses Exploit Code

KEY FACTS

Gemini 4 took part in vulnerability discovery tests across six controlled web application labs.
Gemini 4 outscored GLM5.3 and Opus 5 in the vulnerability discovery phase.
When generating reproducible exploits, Gemini 4's recall dropped by 9 percentage points.
Opus and GLM saw recall drops of only about 1 to 2 percentage points.
Gemini 4 exhibited soft refusals, declining exploit tasks during otherwise normal operation.

KEY DATA

9ppGemini 4 recall drop

1-2ppOpus/GLM recall drop

40%Gemini 4 read tool call share

10%GLM 5.3 read tool call share

PHYSIX OBSERVATION

Gemini 4 leads in vulnerability discovery, but soft refusals make exploit generation unstable, exposing the tension between safety alignment and real-world capability. This is a boon for defenders, but attackers may pivot to other models, and the industry needs to watch model behavior predictability.

Source: Hacker News report