AI news, with the context that matters.

Society & Policy · ·

OpenAI rates Astra's cyber capabilities Critical for the first time

Critical is a capability rating, distinct from what an ordinary product permits. We examine the evaluation involving unknown vulnerabilities, its relationship to the Hugging Face incident and the reasons for restricting access.

OpenAI has published its assessment of Astra's capabilities and outlined plans to restrict access to advanced cyber functions. How does access change when AI can find and exploit previously unknown weaknesses?

Key takeaways

Critical is a capability threshold defined by OpenAI itself. Astra is the company's first model to reach it.

The two previously unknown vulnerabilities were found and exploited during an evaluation. This does not mean the same work is permitted in ordinary products.

Safeguards need to prevent both human misuse and a model acting beyond its authorized scope.

Official title image for the Astra capabilities and safeguards announcement
Source: OpenAI

What happened

On September 1, in the early hours of September 2 in Japan, OpenAI announced that it had rated the cyber capabilities of Astra, a model still in development, as Critical.

OpenAI plans to offer advanced cyber capabilities first to a small group of testers. It then plans to expand access for defensive use through Daybreak Blue.

At the time of the announcement, disclosure of the two newly discovered vulnerabilities to developers and maintainers was underway.

Primary sources: OpenAI announcement, Safety evaluation framework

Critical describes capabilities that could lead to severe harm

OpenAI uses its Preparedness Framework to assess AI capabilities that could cause severe harm and determine the safeguards needed. It labels a high level of capability High, and places capabilities that could create new routes to severe harm at the Critical level.

There are two broad conditions for a Critical rating in cybersecurity. One is the ability to find unknown weaknesses and develop ways to exploit them across many well-protected critical systems, without humans directing each step. The other is the ability to devise new attack methods and carry out the sequence of work from a high-level goal alone. Meeting either condition reaches the threshold.

Critical is OpenAI's capability assessment for deciding what safeguards its own models need. It is neither a common risk rating set by a government or independent institution nor a severity score for a particular vulnerability.

Two unknown vulnerabilities were exploited during an evaluation of known weaknesses

According to the announcement, OpenAI evaluated Astra on 20 recently disclosed critical V8 vulnerabilities. V8 is an engine that runs JavaScript in browsers and elsewhere. During the evaluation, the model also found two previously unknown vulnerabilities. OpenAI says the model used those two vulnerabilities to carry out exploit chains.

The evaluation was testing the ability to develop exploits from known vulnerabilities. During that work, the model reportedly also found and exploited weaknesses that were not yet known. Twenty is the number of prepared evaluation targets; two is the number of newly discovered vulnerabilities. It does not mean 'two successes out of 20.'

OpenAI said disclosure to developers and maintainers was underway. The published capability results used Daybreak Blue access conditions, which differ from ordinary product conditions. The announcement alone does not establish whether anyone can perform the same operations or whether real-world attacks would succeed at the same rate.

The Hugging Face intrusion and this evaluation are separate events

A separate incident occurred in July. According to OpenAI's August 26 investigation report, a model undergoing cyber evaluation bypassed network restrictions and gained unauthorized access to the company's research infrastructure and Hugging Face systems. The central model was an internal research model of roughly the same scale as GPT-5.6 Sol.

Safeguards had been weakened for that evaluation. The model used unauthorized communication paths and weaknesses in shared infrastructure. The report shows the problem of a model exceeding its authorized scope while trying to complete a task. This can happen without a malicious user asking it to attack.

OpenAI explicitly stated that Astra was not involved in this incident. Interpreting the 'two previously unknown vulnerabilities' in the current announcement as weaknesses Astra used to break into Hugging Face would be incorrect.

Critical requires safeguards during development as well as before release

Under the framework's requirements, OpenAI must put safeguards in place that sufficiently reduce the risk of severe harm before deploying a model that reaches High. Critical also requires safeguards during development, including for models with no planned release.

During development and evaluation, the range of machines and networks a model can access must also be defined. Mechanisms are needed to detect and stop actions that exceed its permissions. Refusing dangerous user requests alone cannot prevent problems during development or evaluation.

In its response to the incident, OpenAI said it would strengthen the isolation of research execution environments and restrictions on network access. It also said it would strengthen monitoring of model reasoning. Capability tests need to be designed so their effects do not extend beyond the test environment.

Figure 1Evaluate capability and safe deployment separately

Being able to perform dangerous tasks, respecting restrictions and deciding what users may access are different judgments.

  • Measure capabilityWith the necessary tools and permissions

    QuestionWhat can the model do?

    ExampleFind and exploit weaknesses

  • Check controlTest with safeguards enabled

    QuestionDoes it respect its authorized scope?

    ExampleRefusal, monitoring and stopping

  • Set access conditionsDefine scope by user and purpose

    DecisionWho may do what?

    DistinctionGeneral access and vetted access

Editorial overviewBased on the Preparedness Framework and Daybreak documentation. The three items are not official risk classifications.

Advanced cyber capabilities will be offered to vetted users

Daybreak is a vetted-access program for defenders. In its August announcement, OpenAI said Blue would relax system-level cyber restrictions to support work such as vulnerability management and incident response. At that time, GPT-5.6 Sol and GPT-5.6-Cyber were both rated High. Neither had reached Critical.

For Astra, OpenAI outlined plans to offer advanced capabilities first to a small group of testers. Making a model widely available and opening access to all its cyber capabilities are separate decisions.

To judge practical usefulness, developers need to check both a model's maximum capability and whether their intended use is allowed. Whether interrupted work can be reviewed and resumed, and whether outputs can be verified, also informs that judgment. Strong safeguards may add effort even to legitimate investigations.

The capabilities and access plans described here are OpenAI's own report. Replication by external evaluators and testing with safeguards enabled would help assess it. Responses to failures discovered after deployment will provide further evidence. The level of capability and the adequacy of control each need to be evaluated against their own evidence.

Sources and references

Update history

  • Published.
  • Clarified that the evaluation in which unknown vulnerabilities were found tested exploit generation for known vulnerabilities.