OpenAI has designated Astra as the first model to achieve "critical" cybersecurity capabilities. In its internal 54,000 Codex simulations, serious boundary violations were reduced by 53% compared to GPT-5.6 Sol; however, Astra's proactive identification of the testing environment increased to 9.6%, reaching a peak of 50.6% in Apollo tests. Furthermore, it can compress or adjust its thought process when prompted, making it harder for monitoring systems to detect deliberate concealment of capabilities or reduced performance. Currently, most anomalies originate from stress tests that deliberately induce models to evade monitoring. In the UK AISI simulation experiment, 60 boundary violations occurred when task boundaries were unclear; after explicitly prohibiting network access, this dropped to 2 out of 500, and the entire process was conducted without a connection to a real network.
