Google announces Gemini 4 Carbon cyber results and phased access for defenders

Key facts

As stated in the sources, with where to find them.

• Google labels the three scores as discovery, exploitation and patching. SEC-bench Pro’s authors describe known-vulnerability tasks that give source paths, vulnerability class and expected error while withholding the reference proof of concept and root-cause location. Its score is not a blind unknown-vulnerability discovery rate; Google’s announcement does not establish identical harness or grading settings.Google, State-of-the-art cybersecurity performance; SEC-bench Pro v2, Section III-D and Figure 2
• Google reports 0.0% attack success on the evaluated Gray Swan indirect prompt-injection benchmark. That benchmark result does not establish resistance to all adaptive attacks.Advancing frontier model safety & security
• The announcement places general developer, business and consumer access in the future. It does not establish that every applicant has access or that all product settings match unguarded evaluations.Coming soon to developers, businesses, and consumers

Findings that cite this record

No tracked finding cites this record yet.

Key questions this bears on

• Can prompt injection against AI agents be reliably defended?Not reliably. Bounding what untrusted input can trigger helps, but trust boundaries must survive delegation and tool execution.
• How far can measured AI cyber capability be trusted?As a lower or conditional bound. Scores move with budget, pipeline and contamination, and the same benchmark name can hide different setups.
• Is AI shifting the balance between finding and fixing vulnerabilities?Discovery is ahead. Verification, triage, coordinated disclosure and safe repair remain bottlenecks between AI findings and deployed fixes.
• How are attackers using AI agents in real operations?To run parts of intrusions at scale: vendors report a state-aligned actor and criminal groups using open-source agent frameworks and commercial models.

Sources

• Gemini 4 Carbon: our next frontier modelGoogle DeepMindOct 28, 2026Primary sourceAccessed Oct 30, 2026
• SEC-bench Pro: Can Language Models Solve Long-Horizon Software Security Tasks?Hwiwon Lee et al. (arXiv)Jul 20, 2026Primary sourceAccessed Oct 30, 2026

Related records