LLMs

OpenAI releases GPT-6 Cyber, a specialized model for authorized vulnerability research

GPT-6 Cyber achieves 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, with improved alignment to stay within authorized scope during security testing.

Last verified:

Bottom Line Up Front

According to OpenAI, GPT-6 Cyber represents a substantial advancement in AI-assisted vulnerability research, achieving a 48.5% score on ExploitGym—a 51% relative improvement over GPT-5.6 Cyber’s 32.1%—while maintaining 99.93% alignment to authorized task boundaries. Researchers have already deployed the model to uncover 81 previously unknown high-severity Firefox vulnerabilities and develop 948 Linux privilege escalation exploits, signaling a shift in how frontier models enable large-scale security auditing.

GPT-6 Cyber’s Benchmark Performance

OpenAI reports that GPT-6 Cyber significantly outperforms its predecessor on core cybersecurity tasks. The model achieves 48.5% on ExploitGym, which evaluates whether models can successfully exploit a specified vulnerability in a remote target—up from GPT-5.6 Cyber’s 32.1%. The company also developed a new evaluation for offensive web penetration testing to assess vulnerability discovery capabilities, though details on that benchmark were not provided in the announcement.

Under OpenAI’s Preparedness Framework, GPT-6 Cyber is now classified as “Critical” in cybersecurity, an escalation from the “High” classification for GPT-5.6 Cyber. This reflects both the model’s increased capability and OpenAI’s assessment of the heightened risk it poses if misused.

Real-World Vulnerability Discoveries

The model has already contributed to identifying previously unknown security weaknesses in production software. According to OpenAI, researchers using GPT-6 Cyber developed 948 privilege escalation exploits in Linux and identified 81 previously unknown high-severity vulnerabilities in Firefox, despite both codebases undergoing extensive prior testing with frontier models. OpenAI reports that Mozilla has released fixes for many of the identified Firefox vulnerabilities.

Beyond exploit discovery, the model also assisted researchers in developing candidate patches, integrating both offensive and defensive security workflows into a single tool.

Alignment to Task Boundaries

A distinguishing feature of GPT-6 Cyber is its improved adherence to authorized scope. In OpenAI’s alignment evaluation, the model avoided exploiting out-of-scope services in 99.93% of tests without a deployment prompt, and 100% with the prompt active. This reflects dedicated training to respect restrictions, stay within authorized scope, and report results accurately—addressing a critical concern for tools that can both identify and exploit vulnerabilities.

Institutional Adoption and Expanded Access

According to OpenAI, security teams at organizations including Cisco and CodeRabbit are already using GPT-6 Cyber for vulnerability research and code review. OpenAI has also deployed the model internally to identify and remediate weaknesses in its own infrastructure.

The company indicates plans to expand access through its Daybreak initiative, aiming to bring advanced security research capabilities to organizations of all sizes, independent researchers, and open-source maintainers. OpenAI’s mainline models such as GPT-6 Astra continue to support code review, vulnerability triage, and patch development for broader use cases.

Why This Matters

The release of GPT-6 Cyber reflects a deliberate split between OpenAI’s general-purpose frontier model (GPT-6 Astra) and a specialized variant optimized for authorized security research. This architecture allows OpenAI to deliver substantial improvements on exploit and vulnerability discovery tasks while maintaining alignment guarantees through dedicated training and deployment prompts.

For security teams, the model’s ability to discover previously unknown vulnerabilities in extensively-tested codebases suggests that AI-assisted auditing can uncover gaps that traditional methods or even earlier frontier models miss. The 99.93% boundary-adherence rate in testing—if it holds under real-world deployment conditions—could reduce the friction of deploying such tools in production environments where accidental scope violation carries high consequences.

However, the classification as “Critical” under the Preparedness Framework signals OpenAI’s recognition that GPT-6 Cyber’s capabilities carry elevated dual-use risk. The efficacy of alignment training, deployment prompts, and institutional safeguards will face scrutiny as broader access expands to independent researchers and open-source maintainers operating outside controlled environments. The decision to release capability-specialized models alongside stricter access controls represents a bet that supervised, audited vulnerability research yields larger net security gains than attempting to restrict the capability entirely.

Frequently Asked Questions

How does GPT-6 Cyber perform on exploit development compared to the previous version?

According to OpenAI, GPT-6 Cyber scores 48.5% on ExploitGym, a 51% relative improvement over GPT-5.6 Cyber's 32.1%. ExploitGym tests whether models can successfully exploit a specified vulnerability in a remote target.

What real-world vulnerabilities has GPT-6 Cyber helped identify?

OpenAI reports that researchers using GPT-6 Cyber identified 81 previously unknown high-severity vulnerabilities in Firefox and developed 948 privilege escalation exploits in Linux, despite both codebases having undergone extensive prior testing with frontier models.

How does GPT-6 Cyber avoid exploiting unauthorized targets?

In alignment testing, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without a deployment prompt, and 100% with the prompt active. This reflects dedicated training to stay within authorized scope and respect task restrictions.

Who is currently using GPT-6 Cyber?

According to OpenAI, security teams at organizations including Cisco and CodeRabbit are using GPT-6 Cyber for vulnerability research and code review. OpenAI has also used it internally to find and fix weaknesses in its own infrastructure.

#gpt-6 #cybersecurity #alignment #vulnerability-research