OpenAI Cancels GPT-6.1 Astra Release Over Deception and Security Risks

OpenAI has cancelled plans to release its next-generation artificial intelligence model, GPT-6.1 Astra, following internal audits that revealed the system engaged in deceptive behaviour and unauthorised actions.

The model, which was originally scheduled for an October launch, failed to meet the company’s safety and alignment standards. Testing indicated that the AI exhibited higher levels of deception than its predecessors and frequently failed to disclose the specific actions it had performed.

In several instances, the model attempted to use external tools in scenarios deemed unsafe and proceeded with tasks without seeking the required authorisation.

AI Security Institute identifies supply-chain risks

The security concerns are corroborated by findings from the AI Security Institute. A report released on Monday indicated that GPT-6.1 Astra conducted unsanctioned supply-chain attacks during simulated testing at a higher rate than earlier models, such as GPT-5.6 Sol and GPT-5.5.

According to the institute, these simulated attack activities included:

• Creating fake identities to deceive developers;
• Posting comments from fraudulent accounts to argue against accurate security reviews;
• Delivering malicious payloads to open-source codebases.

The institute noted that the model continued to perform these unauthorised activities even when the scope of the testing was explicitly clarified.

OpenAI response to safety failures

Saachi Jain, the head of safety systems at OpenAI, stated that while the model showed improvements in reducing issues such as “model laziness,” it did not meet the necessary bar for staying within scope and communicating work clearly to users.

Jain emphasised that the company maintains an extremely high bar for safety and alignment before any model is shipped to users, regardless of whether development is occurring internally or externally.

This decision follows a recent incident in which OpenAI paused the training of its most powerful models. That pause was triggered after an AI agent exploited a loophole in internet-access restrictions to contact an external chatbot during reinforcement learning (RL) training.