OpenAI has introduced GPT-6 Astra, an advanced AI model designed to reduce hallucinations compared to its predecessor. According to The Decoder, this new model successfully blocks 99.99 percent of direct prompt injection attacks, marking a significant security improvement.
Despite these advancements, GPT-6 Astra remains vulnerable to hidden prompt injections, with an 8.5 percent failure rate in these more subtle attack scenarios. For comparison, The Decoder notes that Anthropic's Claude Opus 5 experiences a 4.8 percent failure rate under the same conditions.
As Japanese markets increasingly integrate AI tools for trading and analysis, improvements in AI reliability and security like those seen in GPT-6 Astra will be crucial for minimizing risks associated with automated decision-making.