OpenAI bolstered security for GPT-6.2 Astra to stop unauthorized model replication. The company reported that the success rate for extracting protected reasoning fell to 0.2% from 54% in the previous version. OpenAI now uses strict safeguards and API restrictions to block these attacks, which increased significantly over the summer.
OpenAI has bolstered the security of its latest AI model, GPT-6.2 Astra, to thwart unauthorized attempts at extracting internal reasoning and replicating advanced capabilities. The company reported a drastic drop in the success rate of such attacks after implementing these security upgrades. The average success rate for attacks trying to extract protected reasoning from an earlier request has dropped to 0.2% in GPT-6.2 Astra, from 54% in its predecessor version with original safeguards.
Distillation is a machine-learning technique where one AI model learns from the data generated by another.
While it has legitimate uses, unauthorized distillation can let another model replicate advanced capabilities without inheriting the original system's safety protections.
OpenAI warned that this could pose risks in sensitive areas such as cybersecurity and biological research, where capabilities developed for defensive purposes could be weaponized.
OpenAI has focused on protecting its models' chain-of-thought, or the intermediate reasoning steps generated while working toward an answer.
This information can contain details not included in the final response and can provide valuable training material for competing models.
An internal study found that AI models trained on exact reasoning traces generally performed better in cybersecurity benchmarks after supervised fine-tuning, than those trained on reasoning summaries.
To combat the threat of unauthorized distillation, OpenAI is using model-level safeguards, input and output classifiers, API restrictions, and investigations into coordinated extraction campaigns.
The company has detected multiple unauthorized distillation campaigns on its platform this year, with weekly enforcement actions ranging from thousands to tens of thousands of accounts.
The pace of such activity increased significantly over the summer.
OpenAI also shares indicators of compromise with other AI model providers to help identify and disrupt similar activities across platforms.
The company has warned that foreign actors could replicate advanced capabilities developed by US AI companies and deploy them without comparable safeguards.
It has called for stronger coordination among AI developers, deployment partners, and governments to protect frontier models from unauthorized capability extraction as advanced AI systems become more widely available.
