SafetyPublished ExplainedAll news from that day
Ongoing case: Cyber-capable AI models 25 stories
How a system card reports a model's tests, ratings and access limits
OpenAI's deployment safety hub says it shares how deployed models perform in evaluations, the risks measured and the steps taken; the GPT-6 Cyber card of 5 November 2026 shows the layout.
Check our sources · 11 facts from 2 sources
Illustration made with AI for ai notis
Key points
- OpenAI's deployment safety hub says it shares how deployed models perform in evaluations, the risks measured and the steps taken to improve.
- The GPT-6 Cyber card, dated 5 November 2026, has sections on training, model safety, alignment and agent safety, safeguards, and cybersecurity evaluations such as ExploitGym.
- The card gives the designations as Critical in Cybersecurity, High in the Biological and Chemical domain, and below High in AI Self-Improvement, with access through Daybreak Red.
What happened
OpenAI said GPT-6 Cyber is rated Critical in cybersecurity, and the place that rating is stated is its system card, dated 5 November 2026.
OpenAI keeps these cards on a deployment safety hub. The hub describes itself as sharing the technical work OpenAI does to make its systems safe, including how deployed models perform in evaluations, the risks measured and the steps taken to improve. For one other card, on ChatGPT Images 2.6, the hub says it covers capabilities, safety evaluations, safeguards and image provenance.
The GPT-6 Cyber card shows how that works in practice. It opens with an introduction that calls the model a cyber-specialized variant of GPT-6 Astra and summarises its ratings, access controls and safeguards.
It then covers training, model safety (including cyber refusals and safe completions), alignment and agent safety, and safeguards. Preparedness and cybersecurity evaluations follow, with named tests such as ExploitBench, SEC-Bench Pro, SRE-Bench and ExploitGym.
In the section on external evaluations, OpenAI says it worked with NIST CAISSI and Irregular on external assessments, and the card summarises only Irregular's findings.
The ratings come from OpenAI's Preparedness Framework, and the card states them in one line: Critical in Cybersecurity, High in the Biological and Chemical domain, and below High in AI Self-Improvement.
The card also holds the access rules. It says access to GPT-6 Cyber is restricted through Daybreak Red, so only approved users and organisations can use it, and that Daybreak Red API access requires IP allowlisting.
Underlined sentences link to the facts we checked. Dotted words open a short explainer.
Fact 3 of 11 · from OpenAI, 5 Nov 2026
Model Data and Training Model Safety Cyber Refusals Expand Safe Completions Safe Completions Evaluations with Challenging Prompts Agentic Safe Completions Biological Safety and Over-Refusal Alignment and Agent Safety Respecting Restrictions and Scope Avoiding Deceptive Interactions Alignment in Workplace Tasks Alignment on Previous Cases of Harmful Outcomes Safeguards Access to GPT-6 Cyber Existing safeguards New and strengthened safeguards
Fact 9 of 11 · from OpenAI, 5 Nov 2026
Daybreak Red provides cyber-specialized models under stricter eligibility and security requirements. Like past models, GPT-6 Cyber is available only to approved Daybreak Red users and organizations.
Fact 11 of 11 · from OpenAI
System card for ChatGPT Images 2.6, covering model capabilities, safety evaluations, safeguards, and image provenance.
What it means for you
Our viewA system card is the lab's own account of what it tested and what limits apply. Read the GPT-6 Cyber card in its own order: the introduction summarises the ratings, access controls and safeguards, the evaluation sections follow, and the external evaluations section is where outside assessments such as Irregular's appear.
Separate the lab's own results from that outside section when you weigh a claim. Before you approve a model for work, check the date on its card and note which access controls, such as required IP allowlisting, your setup would have to meet.
It adds no new facts.
Your reaction
We count reactions per story and day, never who reacted. The counts help us choose what goes in the monthly issue. If you are signed in, your own page shows yours too.
We count reactions per story and day, never who reacted. They help choose what goes in the monthly issue.
Check our sources
Every sentence above is checked against these 2 sources.
1 GPT-6 Cyber System Card
Open the source archived copy-
OpenAI's GPT-6 Cyber System Card is dated 5 November 2026.
citePublished November 5, 2026
-
The GPT-6 Cyber card's introduction describes GPT-6 Cyber as a cyber-specialized variant of GPT-6 Astra and summarises its capability designations, access controls and safeguards.
citeGPT-6 Cyber is a cyber-specialized variant of GPT-6 Astra designed for authorized security work, including tasks such as vulnerability discovery, exploit development, and web pentesting. The model has specialized post-training that aims to improve cybersecurity capabilities and reduce cyber refusals.
-
The card has sections on model data and training, model safety (including cyber refusals and safe completions), alignment and agent safety, and safeguards (access, existing safeguards, new and strengthened safeguards).
citeModel Data and Training Model Safety Cyber Refusals Expand Safe Completions Safe Completions Evaluations with Challenging Prompts Agentic Safe Completions Biological Safety and Over-Refusal Alignment and Agent Safety Respecting Restrictions and Scope Avoiding Deceptive Interactions Alignment in Workplace Tasks Alignment on Previous Cases of Harmful Outcomes Safeguards Access to GPT-6 Cyber Existing safeguards New and strengthened safeguards
-
The card's cybersecurity evaluations section covers ExploitBench, its internal port, SEC-Bench Pro, SRE-Bench, ExploitGym and ExploitGym Hardened.
citeCybersecurity evaluations ExploitBench ExploitBench — Internal Port (June–August 2026) SEC-Bench Pro SRE-Bench ExploitGym ExploitGym Hardened
-
The card's section on external evaluations for cyber covers NIST CAISSI and Irregular assessments.
citeWe also worked with NIST CAISSI and Irregular on external assessments of GPT-6 Cyber’s cybersecurity capabilities.
-
The card gives the designations as Critical in Cybersecurity, High in the Biological and Chemical domain, and below High in AI Self-Improvement.
citeBased on these evaluations, we give GPT-6 Cyber the same preparedness capability designations as GPT-6 Astra: Critical in Cybersecurity , High in Biological and Chemical capabilities, and below High in AI Self-Improvement
-
The card describes required IP allowlisting for Daybreak Red API access.
citeAPI IP allowlisting. To make stolen credentials harder to use, we require IP allowlisting for Daybreak Red API access by default.
-
The card says that, like GPT-6 Astra, OpenAI classifies GPT-6 Cyber as Critical in Cybersecurity under its Preparedness Framework.
citeLike GPT-6 Astra, we classify GPT-6 Cyber as Critical in Cybersecurity under our Preparedness Framework.
-
The card says access to GPT-6 Cyber is restricted through Daybreak Red, and that only approved Daybreak Red users and organisations can use the model.
citeDaybreak Red provides cyber-specialized models under stricter eligibility and security requirements. Like past models, GPT-6 Cyber is available only to approved Daybreak Red users and organizations.
2 OpenAI deployment safety hub
Open the source archived copy-
OpenAI's deployment safety hub describes itself as covering the technical work OpenAI does to make its systems safe, including how deployed models perform in evaluations, the risks measured and steps taken to improve.
citeSharing the technical work we do to make our systems safe, including how deployed models perform in evaluations, the risks we measure, and the steps we take to improve over time.
-
The hub describes the ChatGPT Images 2.6 system card as covering model capabilities, safety evaluations, safeguards and image provenance.
citeSystem card for ChatGPT Images 2.6, covering model capabilities, safety evaluations, safeguards, and image provenance.
Every sentence above is checked against these 2 sources. Each fact has its own address you can cite.
Source 1 · OpenAI · 5 Nov 2026
GPT-6 Cyber System Card
The GPT-6 Cyber card's introduction describes GPT-6 Cyber as a cyber-specialized variant of GPT-6 Astra and summarises its capability designations, access controls and safeguards.
GPT-6 Cyber is a cyber-specialized variant of GPT-6 Astra designed for authorized security work, including tasks such as vulnerability discovery, exploit development, and web pentesting. The model has specialized post-training that aims to improve cybersecurity capabilities and reduce cyber refusals.
The card has sections on model data and training, model safety (including cyber refusals and safe completions), alignment and agent safety, and safeguards (access, existing safeguards, new and strengthened safeguards).
Model Data and Training Model Safety Cyber Refusals Expand Safe Completions Safe Completions Evaluations with Challenging Prompts Agentic Safe Completions Biological Safety and Over-Refusal Alignment and Agent Safety Respecting Restrictions and Scope Avoiding Deceptive Interactions Alignment in Workplace Tasks Alignment on Previous Cases of Harmful Outcomes Safeguards Access to GPT-6 Cyber Existing safeguards New and strengthened safeguards
Show all 9 factsShow fewer facts
The card's cybersecurity evaluations section covers ExploitBench, its internal port, SEC-Bench Pro, SRE-Bench, ExploitGym and ExploitGym Hardened.
Cybersecurity evaluations ExploitBench ExploitBench — Internal Port (June–August 2026) SEC-Bench Pro SRE-Bench ExploitGym ExploitGym Hardened
The card's section on external evaluations for cyber covers NIST CAISSI and Irregular assessments.
We also worked with NIST CAISSI and Irregular on external assessments of GPT-6 Cyber’s cybersecurity capabilities.
The card gives the designations as Critical in Cybersecurity, High in the Biological and Chemical domain, and below High in AI Self-Improvement.
Based on these evaluations, we give GPT-6 Cyber the same preparedness capability designations as GPT-6 Astra: Critical in Cybersecurity , High in Biological and Chemical capabilities, and below High in AI Self-Improvement
The card describes required IP allowlisting for Daybreak Red API access.
API IP allowlisting. To make stolen credentials harder to use, we require IP allowlisting for Daybreak Red API access by default.
The card says that, like GPT-6 Astra, OpenAI classifies GPT-6 Cyber as Critical in Cybersecurity under its Preparedness Framework.
Like GPT-6 Astra, we classify GPT-6 Cyber as Critical in Cybersecurity under our Preparedness Framework.
The card says access to GPT-6 Cyber is restricted through Daybreak Red, and that only approved Daybreak Red users and organisations can use the model.
Daybreak Red provides cyber-specialized models under stricter eligibility and security requirements. Like past models, GPT-6 Cyber is available only to approved Daybreak Red users and organizations.
Source 2 · OpenAI
OpenAI deployment safety hub
OpenAI's deployment safety hub describes itself as covering the technical work OpenAI does to make its systems safe, including how deployed models perform in evaluations, the risks measured and steps taken to improve.
Sharing the technical work we do to make our systems safe, including how deployed models perform in evaluations, the risks we measure, and the steps we take to improve over time.
The hub describes the ChatGPT Images 2.6 system card as covering model capabilities, safety evaluations, safeguards and image provenance.
System card for ChatGPT Images 2.6, covering model capabilities, safety evaluations, safeguards, and image provenance.
Topics
The morning email
On the mornings we publish: the three top stories and up to four short ones. Free.
We email you a link to confirm. An issue may include one sponsor, always labelled Sponsored · Advertisement. Our emails count opens and clicks, not who made them. Unsubscribe in one click. What we keep