OpenAI Recognized Astra as the First Model With Critical Cybersecurity Capabilities

• Astra can independently discover previously unknown vulnerabilities and build exploit chains in hardened systems.
• In tests, the model uncovered two zero-day vulnerabilities in the V8 engine, and the dangerous cyber request refusal rate rose to 91.5%.

OpenAI said that its Astra artificial intelligence model has reached the threshold for critical cybersecurity capabilities within its preparedness framework.

According to OpenAI, with the right tools and access, Astra can find previously unknown vulnerabilities and develop ways to exploit them in well-protected systems without step-by-step human guidance. Astra is the first OpenAI model to be classified at this level.

The company noted that because of this level of capability, it has delayed certain stages of the model’s development and release over the past few weeks to strengthen and test protections against cyber misuse and unauthorized actions.

What did Astra’s tests show?

Under OpenAI’s preparedness framework, a model reaches the critical threshold if it can:

• find and create functional zero-day exploits for many hardened critical systems without human intervention
• develop and execute end-to-end novel cyberattack strategies against protected targets with only a high-level objective

Astra’s evaluation included automated public and private benchmarks, as well as expert-led experiments.

On the ExploitBench benchmark, the model scored 100% when assessing its ability to develop exploits for known vulnerabilities.

Separately, OpenAI created an internal benchmark, ExploitBench — Internal Port, which includes 20 recently disclosed high-severity vulnerabilities in the V8 engine. Astra demonstrated a significantly higher level of arbitrary code execution than GPT-5.6 Sol, while using far fewer output tokens.

During this evaluation, Astra also discovered and exploited two zero-day vulnerabilities as part of an exploit-chain attack. OpenAI said it is currently disclosing details to the developers of the relevant software.

In experiments with a hardened browser and operating system, Astra also found previously unknown vulnerabilities and combined them into working exploit chains.

According to the company, for models with these capabilities, two key risks need to be controlled:

• first — malicious actors using Astra to develop exploits for unknown flaws in protected critical systems or to carry out complex attacks
• second — the possibility that the model itself could take unauthorized actions due to issues with its alignment.

In addition, OpenAI said it has further strengthened Astra’s model hardening within its safeguards system, and improved handling of cross-session context. In tests, Astra refused 91.5% of dangerous requests, compared with 59% for GPT-5.6 Sol.

For higher-risk accounts, the company applies more conservative limits on model behavior and enhanced monitoring for potential cyber abuse.

OpenAI also continues internal and external red-teaming, regression testing, and the search for new jailbreaks. The company plans to use a 24/7 rapid response program to investigate and remediate new vulnerabilities.

Astra will initially be available to a limited group of users

OpenAI plans to make Astra available in the near future. At the same time, the model’s most advanced cybersecurity capabilities will initially be available only to a limited group of testers.

Cutting-edge cybersecurity work will first be opened to a small group of alpha testers, after which access will be expanded via the Daybreak Blue program for defensive use.

The company warned that the strengthened safeguards may sometimes mistakenly flag legitimate activity as potential cyber abuse or unauthorized behavior. In that case, the task may be slowed, paused, or stopped.

As a reminder, earlier OpenAI and another 155 companies called for stronger AI cybersecurity.

Сообщение OpenAI Recognized Astra as the First Model With Critical Cybersecurity Capabilities появились сначала на INCRYPTED.