OpenAI Strengthens Defenses Against Adversarial Distillation
Reported by OpenAI, Blockchain News and Superpower Daily3 sources ↓
OpenAI says updated safeguards in GPT-6.2 Astra reduced success in internal tests of protected-reasoning extraction from 54.0% to 0.2%.
Key points
- OpenAI says its approach combines model training, safeguards, detection and enforcement
- In internal tests, extraction success fell from 54.0% for GPT-6 Astra to 0.2% for GPT-6.2 Astra
- The protections also apply to GPT-6.2 Sol, but no test results for Sol were reported
Sources
Details
OpenAI’s measures target attempts to use a model’s outputs to train another model that reproduces its capabilities without the same safeguards. The company identifies protected reasoning traces as one type of information that could be extracted for this purpose, including in cybersecurity and biology-related use cases.
OpenAI’s approach includes adversarial testing, model training, classifiers, API-level protections and investigations into coordinated extraction attempts. The reported test compared GPT-6 Astra with its original safeguards against GPT-6.2 Astra with updated protections; the figures are from OpenAI’s internal automated red-team evaluation.
Written from coverage as of