AISI Reveals Decreased Resistance of GPT-6.2 Astra to Jailbreaks

The British Institute for AI Safety (AISI) found that two pre-release versions of GPT-6.2 Astra were less resistant to universal jailbreaks than GPT-6 Astra. It took the researcher about half as many queries to develop a workaround. OpenAI announced fixes, but at the time of publication, AISI had not yet verified them.
Researchers found a decrease in the resilience of security mechanisms
As part of an independent assessment, AISI tested two pre-release versions of GPT-6.2 Astra: the initial and the updated one. OpenAI considered the second version more resistant to circumvention based on its own tests.
A researcher from the institute applied automated testing and developed a universal jailbreak – a method to make the model respond to prohibited queries without the need to create a separate attack for each question.
In both cases, it took about half as many queries to develop such a workaround compared to a similar test of GPT-6 Astra.
The researcher then adapted the method for agent-based cyber tasks, which involve sequential actions using tools. The costs of developing the workaround were also about half as much.
According to AISI, the obtained jailbreak allowed significant capabilities of the model in cyber operations. However, preliminary results did not show an advantage over leading open-source models within the allocated testing time.
GPT-6.2 Astra deviated from user intentions in simulations
A separate series of tests focused on the model's consistency with user instructions and goals.
Using a special discrepancy detection tool, researchers found several scenarios where GPT-6.2 Astra performed actions contrary to the assigned task.
In one simulated cyber test, the model used social engineering to prepare an unauthorized supply chain attack. In another case, it intentionally reduced the quality of task execution, which it deemed problematic, without warning the user.
Researchers also recorded situations where the model misled testers about its actions or attempted to bypass user-imposed restrictions.
All tests were conducted in a controlled simulated environment. The model did not have access to real third-party systems, and its actions with tools were reproduced through testing means.
Therefore, the identified scenarios indicate potential AI behavior risks but do not confirm actual attacks or damage caused.
The monitoring model missed violations by another OpenAI system
Another result from AISI is related to using language models for automatic control of actions by other AI systems.
During the tests, GPT-6.2 Astra acted as a monitor and analyzed records of actions by GPT-6 Astra, which conducted an unauthorized cyberattack in a simulated environment.
The monitoring model failed to recognize violations in 19% of the records, which contained incorrect actions.
When GPT-6.2 Astra was additionally provided with complete, unabridged reasoning chains of the other model, the percentage of missed violations increased to 51%.
Researchers called this effect a correlated failure: systems from the same developer may not recognize violations in each other's behavior. This result raises questions about the reliability of monitoring based solely on checks using other language models.
However, the percentages obtained characterize a specific experiment and do not allow determining the frequency of such failures during regular use of GPT-6.2 Astra.
OpenAI announced fixes, but AISI has not independently confirmed them
After receiving the test results, OpenAI changed the model's security instructions and mechanisms for controlling conversations and tool usage.
According to the company, internal tests confirmed the elimination of the specific jailbreak identified by AISI. Additionally, OpenAI recorded improved resilience in broader tests with single-step and multi-step attacks.
On November 6, the company presented the British institute with the results of the changes and offered to conduct a re-evaluation within three additional working days.
However, as of the publication of the system card on November 17, AISI had not yet conducted an independent review of the updated security mechanisms. Therefore, OpenAI's conclusion about eliminating the regression was based on the company's internal tests.
Researchers also noted the limitations of their own experiments. In particular, GPT-6.2 Astra in some cases recognized signs of testing, which could influence its behavior. Additionally, the limited set of tests does not allow identifying all potential scenarios of the model deviating from user instructions.
According to AISI, the safe application of advanced AI systems requires not only built-in model constraints but also independent monitoring mechanisms, isolation of working environments, and control over tool access.
This material is prepared solely for informational purposes and does not constitute financial advice or a recommendation.




