共有:

Critical Cyber Tier

An AI just got
filed under"Critical."

OpenAI has classified its newest model, GPT-6 Astra, at the top of its four-tier safety framework for cybersecurity risk — the first model ever to reach that ceiling.

AI Navigate Editorial2026.09.227 min read

LOW MEDIUM HIGH CRITICAL GPT-6 Astra Preparedness Framework — Cybersecurity
01

What Happened

An AI that finds flaws
and writes the exploit

In September 2026, OpenAI disclosed on its official blog that its latest model, GPT-6 Astra, had reached "Critical" — the highest of the four cybersecurity risk tiers defined by its Preparedness Framework — for the first time. The framework's top tier had until now existed mostly as a theoretical ceiling; this is the first case of a model actually reaching it.

According to OpenAI's own description, given the right tools and access, Astra can find previously unknown vulnerabilities across many well-defended systems and build new exploits for them without a person guiding each step. It is, by OpenAI's own account, capable of both discovering and weaponizing flaws with minimal human involvement.

Previous modelsGPT-6 Astra
Mostly explained/classified known flawsAutonomously discovers unknown flaws
Exploit code assumed human assistanceDiscovery through exploit-building in one pass
Risk tier: High or belowTier: Critical (top)
Shipped open to everyone at GAAdmin opt-in required per workspace
02

What the evaluations showed

100%
ExploitBench: turned known flaws into exploits
39%
Success rate on vulnerabilities from the last 3 months
2
Zero-days found during pre-release testing

Per OpenAI's "Path to Astra" safety report, Astra scored 100% on ExploitBench — a benchmark that measures whether a model can turn documented vulnerability disclosures into working exploit code. It also succeeded on 39% of novel vulnerabilities disclosed within the prior three months, i.e. flaws that were unlikely to be in its training data. During pre-release evaluation, Astra itself discovered two zero-day vulnerabilities — flaws nobody had previously identified.

OpenAI says these results led it to delay the originally planned release and add further safeguards before shipping.

01

Release delayed

Once the Critical rating was confirmed, OpenAI paused its planned launch schedule to spend more time on additional safety testing and mitigation design.

02

Off by default

Even after general availability, Astra stays disabled by default in enterprise workspaces until an administrator manually turns it on from the admin console.

03

Staged rollout

It first ships to a limited set of organizations, then expands to ChatGPT Plus/Pro/Business/Enterprise, the API, and AWS in stages.


A blade that cuts both ways for defense and offense — companies now have to
draw it on their own judgment.


03

Who It Affects

Who actually feels this

A Critical rating isn't a "don't use this" warning — it's closer to "use it, but govern it." How much it matters depends on where you sit.

Security teams

It could be a powerful ally for automating vulnerability scans and pentests, but the same capability now demands a fresh threat assessment for what happens if it leaks or gets misused.

Admins and leadership

Turning Astra on becomes an explicit opt-in decision. Before enabling it, check whether your access controls and audit logging are actually ready for it.

Everyday ChatGPT users

Nothing changes for casual use — Astra stays off by default, so day-to-day Q&A and coding help are unaffected.

04

What's Next / Risk

There's reason not to relax

Safety frameworks like this have existed at most labs for a while, but they'd mostly functioned as a hypothetical guardrail nobody expected to actually trip. Astra breaks that pattern — it's the first real case of the framework's top tier firing in practice. The next thing to watch is whether other labs, such as Anthropic and Google DeepMind, apply the same bar when their own frontier models reach comparable capability, since whether a "Critical" label gets attached will shape how enterprises decide to adopt them.

There are limits worth naming, too. This evaluation is OpenAI's own; it isn't independently audited by a third party. "Off by default" is an operational setting, not a technical lock — it can be flipped if an admin account is compromised. And the very capability to find flaws and write exploits cuts both ways: if the model is ever jailbroken, these same benchmark numbers describe what an attacker could do with it. Anyone evaluating adoption should read the delay and the opt-in rollout as a signal of how seriously OpenAI itself is treating the risk.