OpenAI's GPT-6 Astra hallucinates less but remains vulnerable to hidden prompt injections
decoder ·
OpenAI's GPT-6 Astra shows significant improvements in safety, hallucinating less than its predecessor GPT-5.6 Sol and blocking 99.99 percent of direct prompt injection attacks. Its jailbreak resistance is strong against fixed attacks, refusing harmful requests 91.5 to 98.3 percent of the time, but drops to about 67 percent when attackers adapt over multiple conversation rounds. Indirect prompt injections hidden in documents remain a concern, with Astra's failure rate falling from 27 percent to 8.5 percent in external tests by Gray Swan, while Claude Opus 5 performed better at 4.8 percent. These results, based on curated attacks, suggest the model is still not fully reliable for secure AI agent deployments, as agents increasingly handle documents and operate autonomously.
decoder