The BaseOfficialOpenAI
GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents
OpenAI announces GPT-6 Cyber, a security model it rates Critical in cybersecurity under its Preparedness Framework, and reports its own benchmark results, vulnerabilities found in Firefox and Linux, an out-of-scope behaviour test and new misuse safeguards.
Why it matters
The same thing, for different readers.
What it means for you
OpenAI says a new AI model found and exploited many flaws in software such as Firefox and Linux, which it says helps defenders but raises the risk if misused.
For lawmakers
OpenAI classifies its own model as Critical for cybersecurity and limits access to approved customers, an example of a company setting access rules that no regulator required.
For researchers
OpenAI reports a test, informed by the Hugging Face incident, of whether agents attack easy out-of-scope services; it says GPT-6 Cyber avoided them in 99.93% of cases without deployment prompts.
The question it raises
OpenAI says its model's reasoning shows it choosing to stay in scope, raising the question of how far an agent's written reasoning can be trusted as evidence of its conduct.
In their own words
Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.
From the source’s own text.
“On ExploitGym, which tests whether models can exploit a specified vulnerability, GPT‑6 Cyber scores 48.5%, up from 32.1% for GPT‑5.6 Cyber.”
“Based on these results, we classify the model as Critical in cybersecurity under our Preparedness Framework”
“In controlled red-teaming campaigns, agents using GPT‑6 Cyber and other OpenAI models chained vulnerabilities to escape microVMs and execute code on the underlying hosts.”
“As models become more capable and uncover more security vulnerabilities, we are also seeing more attempts to use them to cause harm.”
“The skill floor is dropping for the reconnaissance and vulnerability research needed before an attack.”
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- OpenAI's own evaluations of GPT-6 Cyber against GPT-5.6 Cyber and GPT-6 Astra, including ExploitGym, a new RemoteExploitBench, a web penetration test and an ExploitGym Honeypot test of staying in scope, plus a lab evaluation with Dragos.
- What it found
- GPT-6 Cyber scored 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber, avoided out-of-scope services in 99.93% of honeypot tests without deployment prompts (94.49% for GPT-5.6 Cyber), and developed working exploits for 948 of 1,724 reproduced Linux vulnerabilities.
- What it does not show
- Results are OpenAI's own and are not independently checked here; the full results sit in a system card that was not read. OpenAI notes higher safeguard block rates are not always better.
- Status
- The company’s own report about its own work
The record
- Questions
- How is AI used in attacks, defence and war? · Are AI companies putting stronger safeguards into practice? · Can AI systems act beyond what people intended or allowed?
- People
- Rob T. Lee (Speaks, Chief AI Officer, Dragos) · Brian Grinstead (Speaks, Distinguished Engineer, Firefox)
- Also covered by
- OpenAI (YouTube)
- What we read
- We read the whole source.
- Found by
- The Radar, run of November 6, 2026, via
openai-news(tier 1: the source’s own channel or an official record) - Cite this
- OpenAI, “GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents”, November 5, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/openai-2026-11-05-gpt-6-cyber-aligned-agents