An independent audit of the pre-release version of the flagship language model GPT-6.2 Astra recorded a notable decline in resistance to jailbreak attacks. This is evidenced by the results of testing conducted by the UK AI Safety Institute (AISI). The conclusion is unequivocal: compared to the previous iteration, the model demonstrates a regression in its protective mechanisms rather than a strengthening of them.

Jailbreak attacks are methods of bypassing a model's built-in restrictions that force it to generate content the developer deliberately blocks. It is precisely this attack vector that has become Astra's weak spot: where the previous version withstood the pressure, the new model succumbs to manipulation noticeably more often.

It is telling that this concerns a pre-release build — that is, a version that has not yet been made widely available. This gives the development team a window to refine the protective perimeter before the public launch. However, the very fact of the regression signals a systemic problem: the race to expand functionality and grow the model's capabilities often comes at the expense of safety.

For the market, this is an important precedent. As more and more corporate and institutional clients integrate large language models into critical business processes, resistance to jailbreaks is turning from a technical detail into a factor of reputational and regulatory risk. Independent audits like the AISI check are becoming a de facto mandatory stage in a model's lifecycle — on par with performance benchmarks.

Particular attention should be paid to the fact that the regression was identified specifically in the pre-release version. This underscores the value of external testing: the developer's internal procedures may have failed to catch the degradation of protective mechanisms, whereas a third-party audit recorded it before the product reached the market.

Cryptalist Analysis: The identified regression points to a fundamental trade-off between expanding the model's capabilities and maintaining safety — and this balance is becoming a key evaluation criterion for institutional players. Cryptalist views such incidents as a signal to reconsider approaches to auditing AI systems: resistance to jailbreaks should be assessed as a standalone quality metric alongside performance. For the market, this reinforces the demand for independent verification of models before their commercial launch.