OpenAI Pauses Development of Cyber-Capable Model Over Security Concerns
OpenAI has paused internal work on its Astra model, citing 'critical' cyber capabilities that could pose a risk. The company is scaling up security and testing continues with government agencies and safety groups. The move aims to ensure the model is shared safely with defenders to close security holes before attackers exploit them.
Key points
- OpenAI has identified its Astra model as having 'critical' cyber capabilities, the top rung of its Preparedness Framework.
- The company is pausing internal work on the model and scaling up security measures to mitigate potential risks.
- Government agencies and safety groups will help test the model to ensure its safe use.
- OpenAI aims to share the model with defenders to help close security holes before attackers exploit them.
- The company's safety researcher, Boaz Barak, says the decision to err on the side of caution is 'proud' and necessary.
OpenAI Pauses Development of Cyber-Capable Model Over Security Concerns
OpenAI has made the decision to pause internal work on its Astra model, citing 'critical' cyber capabilities that could pose a risk. The company is scaling up security measures and testing continues with government agencies and safety groups.
Background
OpenAI's Astra model has shown strong results in its testing phase, with the company unable to rule out its potential for critical cyber capabilities. This is the top rung of OpenAI's Preparedness Framework, which was first written in 2023. A model reaches this level if it can find and build working zero-day exploits against many hardened systems with no human help.
Response
OpenAI is taking a measured approach to the situation, isolating test environments and restricting the model's network and tool access. The company is also hardening how it stores the weights and monitoring every agentic run for risky behavior. Government agencies and safety groups will help test the model to ensure its safe use.
Rationale
OpenAI's safety researcher, Boaz Barak, framed the aim of the decision as sharing Astra with defenders safely. The company's bet is that cyber-capable models should help defenders close security holes before attackers reach them. The decision to err on the side of caution is seen as 'proud' and necessary by Barak.
Next Steps
The testing of the Astra model will continue with government agencies and safety groups. The outcome of this testing will determine the next steps for OpenAI and the wider AI community.
Sources
The WireByte editorial team synthesises technology news from multiple primary sources, verifies the facts, and links every source. Articles are produced with AI assistance and reviewed under our editorial policy.