OpenAI's chief scientist and CEO sound 'unmonitorability' warning as Astra reaches Critical cybersecurity threshold
Key Takeaways
- •OpenAI's Astra model is the first AI model to reach the "Critical" cybersecurity designation under the Preparedness Framework, introduced in December 2023.
- •Astra scored 100% on ExploitBench and, in tests on 20 high-severity V8 flaws, discovered two zero-days and outperformed Sol in generating arbitrary code execution.
- •Chief scientist Jakub Pachocki and CEO Sam Altman publicly urged calm, asserting that safety measures and chain-of-thought monitoring are advancing alongside the model's capabilities.
- •OpenAI paused reinforcement-learning training for deployment-bound models from August 18 to August 28 to reinforce research clusters and optimize monitoring, and said Astra was not involved in the Hugging Face incident.
- •Astra's most advanced functions will be limited to members of OpenAI's Daybreak coalition, mirroring Anthropic's restricted access to Mythos 5.1 via Project Glasswing, while most capability and safety claims remain without independent third-party validation.

OpenAI chief scientist Jakub Pachocki spoke out within hours of the company becoming the first in the AI race to reach the "Critical" cybersecurity designation under its Preparedness Framework, following a preview of its upcoming Astra model.
His warning early on Wednesday was intended to cool the hysteria surrounding growing commentary suggesting that the AI lab may no longer have a hedge around the capabilities of its largest models.
Why is OpenAI's Astra 'Critical'?
The Critical designation is a level within the Preparedness Framework, which OpenAI introduced to the AI sector in December 2023. The framework measures how much human prompting and intervention an AI model needs to find and build viable zero-day exploits across hardened systems, or to run a full novel attack.
A zero-day exploit refers to the exploitation of flaws that even the software's makers have not yet discovered. These flaws are among the most prized tools in offensive cybersecurity because no patch exists until the vendor finds and fixes them, which is why AI models capable of finding them autonomously raise distinct security concerns.
OpenAI stated on September 1 in its Path to Astra post that the model will be able to complete all the steps for full novel attacks or zero-day exploits with minimal human intervention when it launches.
As evidence, OpenAI cited Astra's 100% score on ExploitBench, a test of how models can exploit known bugs.
Astra is not only better at plotting zero-day attacks, it is also more efficient. OpenAI demonstrated this using an internal test on 20 high-severity V8 flaws that only recently became public knowledge. V8 is the JavaScript engine used in Google's Chrome browser, one of the most widely deployed pieces of software in the world, which makes flaws in it a frequent target of both security researchers and attackers. Astra consumed fewer tokens, discovered two zero-days, and outperformed Sol at generating arbitrary code execution overall.
The same post said Astra could chain unknown browser bugs into a sandbox escape and stitched together an operating-system privilege escalation from ordinary user to root during expert-led trials.
As TechCrunch noted, Astra joins Anthropic's Mythos model, released earlier this year, as the first large language models with those capabilities. The fact that two rival labs have crossed this capability line within the same year is itself part of what has fueled debate over how quickly frontier cyber capabilities are arriving relative to the defenses around them.
Pachocki calls for calm over Astra
Pachocki is pushing back against the hysteria that OpenAI is entering uncharted territory with Astra. According to him, he would not want "a race into unmonitorability kicked off by confused reporting."
The OpenAI executive insisted the lab did its due diligence on chain-of-thought monitoring, the practice of reading the step-by-step reasoning a model exposes as it works. In the X post, he cited the depth of the computation graph behind OpenAI's current frontier models, Astra included, which sits within a factor of two of GPT-4.
Before Pachocki, OpenAI CEO Sam Altman had also called for calm on the same day the Astra blog was released.
"There is an obvious tension here," he wrote on Tuesday, calling Astra "very good" while arguing that safeguards must advance in step with raw capability. Altman said OpenAI had spent the entire summer doing its part to ensure the model ships with the required safety measures.
The safety discussion has been ongoing for weeks. On August 18, OpenAI committed to pausing reinforcement-learning training for deployment-bound models for two weeks so its engineers could reinforce research clusters and optimize monitoring. Training resumed on August 28.
According to OpenAI, Astra was not involved in the now-infamous Hugging Face incident. The company also said the model did not even attempt to break out of its sandbox, even when presented with a scenario built to tempt it into reenacting the Hugging Face episode.
AI labs are restricting access to frontier models
OpenAI said not everyone will have access to Astra's most advanced functions, reserving those capabilities for organizations in OpenAI's Daybreak coalition, while Daybreak Blue firms will have access to the sharpest defensive features. The tiered-access approach mirrors a broader pattern across frontier AI labs, which have increasingly gated their most capable models behind vetted enterprise or government programs rather than releasing them to all users.
As Cryptopolitan reported, Anthropic also said only firms in its Project Glasswing will have access to Mythos 5.1, which is essentially Fable 5.1 with separate safeguards.
OpenAI also said it will monitor and throttle answers to queries from accounts it judges as higher-risk. Overall, Astra will ship with additional chain-of-thought monitoring to catch bad behavior and refuse harmful cyber requests more often.
Most of the information about Astra was provided by OpenAI itself, with limited third-party validation of the company's safety or capability claims.
Yona Shavit, a former OpenAI employee now working on AI resilience at the OpenAI Foundation, publicly asked whether Astra behaving during tests reflects genuine alignment or a model that simply knew what evaluators wanted to see.
OpenAI says a full system card and further evaluations will arrive when Astra launches widely. Until then, the reach of its cyber skills and the strength of the fences around them remain hard to judge.