AgentsSeptember 4, 2026via The Decoder
OpenAI's GPT-6 Astra hallucinates less but remains vulnerable to hidden prompt injections
Why it matters
Model capability improvements (hallucination reduction) matter less than the security gap that remains for autonomous systems. Hidden prompt injection vulnerability is a critical blocker for agent deployment at scale.
Key signals
- GPT-6 Astra blocks 99.99% of direct prompt injections
- Hidden prompt injections embedded in documents succeed 8.5% of the time in GPT-6 Astra
- Claude Opus 5 performs better at 4.8% vulnerability rate to hidden injections
- Hallucination reduction noted but secondary to security findings
- Direct implication: agents reading untrusted documents remain vulnerable
The hook
GPT-6 Astra blocks 99.99% of direct prompt injections. Hidden ones still crack it 8.5% of the time. For agents handling real data, that's a production risk.
OpenAI's GPT-6 Astra hallucinates less than its predecessor and blocks 99.99 percent of direct prompt injections. But when attacks are hidden inside documents the AI reads, the model still gets cracked in 8.5 percent of scenarios. Claude Opus 5 does better at 4.8 percent. For autonomous AI agents ha…