Breaking

OpenAI’s GPT-6 Astra Scores 100% on ExploitBench as Cybersecurity Safeguards Tighten

September 4, 2026 at 02:50 EDT

Foundation Models
AI Agents
Security

OpenAI has released GPT-6 Astra, reporting a perfect 100% score on ExploitBench and saying the model discovered two previously unknown zero-day vulnerabilities during separate exploit-development evaluations. The public version is restricted to defensive work such as secure code review and patching, and refuses requests to create proof-of-concept exploits.

September 3, 2026 · Cybersecurity

A perfect exploit score, with the brakes on

GPT-6 Astra reached 100% on ExploitBench and found two previously unknown zero-days during separate evaluations. The released version is limited to defensive security work.

ExploitBench

100%

reported score

Recent-vulnerability test

39.0%

arbitrary code execution

Discovery

previously unknown zero-days

ExploitBench performance

GPT-6 Astra

100%

GPT-5.6 Sol

78.5%

Each block represents roughly 10 percentage points; the labels show the reported scores.

Defensive upside

Secure code review, patching, vulnerability analysis and detection engineering can move faster.

Deployment constraint

The released model refuses requests to create proof-of-concept exploits.

The central tension

More autonomous vulnerability research increases both defensive potential and the need for controls.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.