OpenAI to limit access to Astra's most powerful cyber capabilities
A broad release is coming soon, but only a limited group of testers will be able to tap its full power.
TL;DR
- OpenAI is releasing its new model, Astra, soon.
- Astra's advanced cybersecurity features will initially be limited to a small group of testers.
- Astra is the first model designated by OpenAI to reach its 'Critical' cybersecurity capability threshold.
- Safety work on Astra aims to prevent malicious use and unauthorized actions by the model.
- OpenAI acknowledges that safeguards may mistakenly flag legitimate activities, potentially slowing or stopping tasks.
- Astra discovered and chained together two zero-day vulnerabilities during testing.
- OpenAI is disclosing these vulnerabilities to the maintainers.
- The company has implemented stronger safeguards for Astra, including improved refusal of harmful requests and monitoring for unauthorized activity.
- Limiting Astra's cybersecurity capabilities may also hinder legitimate vulnerability research.