#openaisaysastrafindsflawsautonomously ASTRA CAN FIND AND EXPLOIT ZERO-DAY VULNERABILITIES. OpenAI says Astra is its first model to reach the “Critical” cybersecurity capability threshold. During testing, Astra: • Scored 100% on ExploitBench, which tests known vulnerabilities • Discovered and used two zero-days in an exploit chain • Escaped a hardened browser sandbox • Gained root access on a hardened operating system On OpenAI’s newer internal benchmark, Astra achieved much higher code-execution rates than GPT-5.6 Sol while using far fewer output tokens. These results used Daybreak Blue access, not Astra’s default configuration. OpenAI says Astra will be available soon, but advanced cybersecurity access will initially be limited. Follow me on X for more AI stuff $T $FF $ACE