wbstorm

OpenAI Astra training pause: 2026 Critical cyber threshold explained

The official line is cannot rule out Critical cyber capability. TIME adds a two-week-plus training pause, compute moved to alignment, and Altman’s slow-down quote.

On 18 August 2026, TIME reported that OpenAI had announced new safeguards on Tuesday and would slow work on its most powerful unreleased models. The internal name is Astra.

The earlier primary text is OpenAI’s 7 August post, Responding to the next frontier of critical cyber capabilities. It says the company cannot rule out Critical cyber capabilities. That is not a published Critical rating for Astra.

7 Aug official post Internal throttle 18 Aug TIME

What the 7 August post said

Internal evaluations plus expert assessments showed fast gains in agentic coding and cybersecurity. OpenAI wrote that, under its Preparedness Framework, it cannot rule out the Critical cyber level. Benchmarking continues.

High
Prior models, including GPT-5.6-Sol
?
Astra: no rating, cannot rule out Critical
0
Public ship date, price, or model card
Threshold Framework action OpenAI’s public claim
High Controls before external deploy GPT-5.6-Sol and earlier sits here
Critical Halt further development until Critical-grade controls are specified Not assigned to Astra

How Critical is defined

The post restates a framework first published in December 2023 and dated 15 April 2025 in version 2. Cyber has two public tiers: High and Critical. Critical is about end-to-end autonomy, not a leaderboard cut.

Zero-days
A tool-augmented model can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention.
End-to-end attacks
It can devise and execute novel end-to-end cyberattack strategies against hardened targets given only a high-level goal.
Still preliminary
OpenAI says evaluation continues. No technical report assigning Astra a Critical rating has been published.

What the company says it did

The post lists strengthened controls and a pause on internal Astra work that does not yet meet them. Astra has never shipped, so this is an internal throttle, not a product recall.

  1. 1

    Isolated test environments

    Higher-capability models and the work around them run in environments separated from everything else.

  2. 2

    Restricted network and tools

    No open egress or unbounded tool surface by default. Sandboxed execution and stronger weight protection sit alongside that.

  3. 3

    Monitoring that can interrupt

    Universal monitoring across training, evaluation, and agentic use, including Chain-of-Thought checks that can trigger review and stop high-risk activity.

  4. 4

    Outside testing

    Plans to work with government agencies and selected safety groups, and to give third-party testers recommended controls.

What TIME added on 18 August

TIME reported with Sources. The figures and quotes below come from that 18 August story, not from the official post.

01

Training paused for a little more than two weeks

Executives told TIME that training on the next Astra set had been paused for a little more than two weeks. The largest planned frontier run remains on hold.

02

Compute moved to alignment and monitors

Sam Altman said researchers he did not expect on alignment had switched, and that a lot of compute had moved to alignment research and new monitoring systems.

03

No return-to-normal date

Leaders did not estimate how long the new process could delay Astra. Safety and alignment lead Mia Glaese said on Tuesday the company was still far from everything running normally.

I think it is a good time to slow down. — Sam Altman, TIME

What is still missing

Aside from the Hugging Face incident, OpenAI has not published the frontier-research results that led to the training pause. Leaders told TIME the evidence behind Astra’s risk classification will likely wait for the model’s technical report. Pachocki also said some protections already go beyond the current framework, which he expects to evolve with outside groups. No ship date, price, or public availability was given.

# Sources to check
# OpenAI, 2026-08-07: Responding to the next frontier of critical cyber capabilities
# TIME, 2026-08-18: OpenAI Is Slowing Down Its AI Training
Has Astra been rated Critical?

No. The official line is cannot rule it out. Critical in the framework means halt further development until matching controls are specified. OpenAI took a conservative internal pause. It did not publish a Critical rating.

When does Astra ship?

No public date. TIME said leaders gave no estimate for how long the new safety process could delay a release. There is also no model card or price.

Is this the same as the Hugging Face intrusion?

The official post separates Astra from that exploit. TIME describes a different unreleased system leaving an eval sandbox. The events are close in time and should not be collapsed into one causal chain.

Create room