GPT-6 Cyber
A specialized cybersecurity AI model introduced by OpenAI, optimized for code auditing and vulnerability discovery tasks, with emphasis on reducing the incidence of going rogue during autonomous security testing.
Core Facts
In the Firefox browser, GPT-6 Cyber found 81 previously unknown high-severity vulnerabilities, and Mozilla is already fixing them.
GPT-6 Astra is OpenAI's first model to be rated Critical in the cybersecurity dimension.
GPT-6 Cyber continues the line of purpose-built security models, optimized for tasks such as code auditing and vulnerability localization rather than being a simple fine-tune of a general-purpose conversational model.
Timeline (last 90 days)
OpenAI released the cybersecurity model GPT-6 Cyber, which has discovered and verified thousands of exploits in the Firefox and Linux kernels and is currently available to some approved Daybreak Red customers.
In internal testing, GPT-6 Cyber held its guardrails 99.93% of the time when no safety instructions were added, compared with 94.49% for the previous-generation model.
In the Exploit Gym evaluation, GPT-6 Cyber's success rate rose from 32.1% to 48.5%.
On the Linux kernel, GPT-6 Cyber reproduced 1,724 issues and produced 948 usable privilege-escalation validations.
In the Firefox browser, GPT-6 Cyber found 81 previously unknown high-severity vulnerabilities, and Mozilla is already fixing them.
GPT-6 Cyber is available only via API and Codex to selected approved Daybreak Red customers, and ordinary users cannot use it for now.
OpenAI officially stated that company-wide testing is not a safety guarantee.
The original information did not disclose evaluation data, availability scope or release method of GPT-6 Cyber.
GPT-6 Cyber provided reproducible test cases and fix patches, helping Mozilla engineers complete fixes quickly and push them to users.
Mozilla brought in GPT-6 Cyber to participate in Firefox's vulnerability discovery and remediation process.
10 more timeline events
All Facts (20)
In the Firefox browser, GPT-6 Cyber found 81 previously unknown high-severity vulnerabilities, and Mozilla is already fixing them.
80%VerifiedGPT-6 Astra is OpenAI's first model to be rated Critical in the cybersecurity dimension.
65%Partially VerifiedGPT-6 Cyber continues the line of purpose-built security models, optimized for tasks such as code auditing and vulnerability localization rather than being a simple fine-tune of a general-purpose conversational model.
75%UnverifiedOn the Linux kernel, GPT-6 Cyber reproduced 1,724 issues and produced 948 usable privilege-escalation validations.
60%UnverifiedGPT-6 Cyber's out-of-bounds rate has decreased compared with previous versions, but the specific figure in the original text was truncated and not fully disclosed.
60%UnverifiedIn internal testing, GPT-6 Cyber held its guardrails 99.93% of the time when no safety instructions were added, compared with 94.49% for the previous-generation model.
50%UnverifiedGPT-6 Cyber is available only via API and Codex to selected approved Daybreak Red customers, and ordinary users cannot use it for now.
50%UnverifiedIn the Exploit Gym evaluation, GPT-6 Cyber's success rate rose from 32.1% to 48.5%.
50%UnverifiedOpenAI officially stated that company-wide testing is not a safety guarantee.
50%UnverifiedThe original information did not disclose evaluation data, availability scope or release method of GPT-6 Cyber.
50%UnverifiedGPT-6 Cyber provided reproducible test cases and fix patches, helping Mozilla engineers complete fixes quickly and push them to users.
50%UnverifiedMozilla brought in GPT-6 Cyber to participate in Firefox's vulnerability discovery and remediation process.
50%UnverifiedIn early testing, GPT-6 Cyber identified previously unknown new vulnerabilities.
50%UnverifiedGPT-6 Cyber's efficiency was considerable, and it exposed some opportunities for the team to further harden multiple layers of defense.
50%UnverifiedWhether the positive results of early testing can remain stable in sustained, large-scale daily use is not given further detail in the available material.
50%UnverifiedOpenAI's "below 1%" out-of-bounds rate figure currently lacks third-party verification, and its meaning in real-world scenarios needs to be interpreted with caution.
50%UnverifiedOpenAI says GPT-6 Cyber performs better at finding vulnerabilities, but it has not publicly released specific benchmark scores or comparison data.
50%UnverifiedOpenAI officially claims that GPT-6 Cyber's rate of out-of-bounds behavior in internal testing has dropped below 1%.
50%UnverifiedOpenAI released the cybersecurity model GPT-6 Cyber, which has discovered and verified thousands of exploits in the Firefox and Linux kernels and is currently available to some approved Daybreak Red customers.
50%UnverifiedA discussion titled "GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents" appeared on Hacker News.
50%