Sakhanda Wire
NVDA $153.61 -1.62% MSFT $506.27 -2.26% GOOGL $280.41 -1.17% META $623.08 -0.53% AMZN $257.66 -0.03%
← Back to the news

GPT-6 Astra reduces hallucinations but remains exposed to hidden attacks

OpenAI’s GPT-6 Astra makes fewer factual errors than GPT-5.6 Sol and performs better against several prompt-based attacks, according to the company’s system card. The hallucination tests used ChatGPT conversations previously flagged by users for incorrect answers, so OpenAI says the results represent especially difficult cases rather than typical everyday performance. The largest improvements appeared with low latency and lower reasoning settings.

Astra blocked 99.99 percent of direct prompt-injection attempts, which target the model through a user’s own instructions. OpenAI attributed the result to GPT-Red, an automated attack method used during training. On a fixed jailbreak dataset involving biology, violence and cybersecurity, the model refused harmful assistance in 91.5 to 98.3 percent of cases. With adaptive attacks across multiple conversation rounds, its defense rate fell to about 67 percent. These tests used the base model without production safety layers.

Indirect injections embedded in documents remained a weakness. Gray Swan’s IPI Arena evaluation used 1,810 attacks and 15 attempts per scenario; Astra was compromised in 8.5 percent of scenarios, compared with 27 percent for GPT-5.6 Sol and 4.8 percent for Claude Opus 5. The findings highlight continuing risks for agents that read documents, use tools or operate computers autonomously.

Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.

Originally published by The Decoder on

Read the original on The Decoder ↗

The reporting and the image belong to The Decoder and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news