2026-11-20 · ← News
OpenAI and Anthropic are closing a shortcut that may have fed open-weight models
OpenAI cut the success rate of a protected-reasoning attack from 54.0% to 0.2% in an internal test, while Anthropic reports a 90 to 99% drop in harvested tokens. Ethan Mollick thinks this may slow open-weight progress, but that remains a hypothesis rather than a measured outcome.
The image could not be loaded.
OpenAI and Anthropic have tightened defenses against unauthorized distillation, in which another team turns a powerful model's outputs into training data. Ethan Mollick argues that the loss of easy access to those data may explain slower open-weight progress. There is no direct evidence for that causal link yet.
Two labs sharply reduced leakage of protected reasoning
OpenAI says an adaptive black-box evaluation cut the success rate for extracting earlier protected reasoning from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra. Extraction from the current request fell from 22.9% to 0.9%. The test aggregated six attacker models and was conducted by the company itself.
Anthropic describes a two-stage classifier, restrictions on moving reasoning across accounts and models, and faster disruption of fraudulent account networks. It estimates that extracted reasoning tokens on its platform fell by a factor of 10 to 100. It also acknowledges a false-positive rate below 0.1% and integration changes for some legitimate customers.
Low-cost followers may lose a valuable source of training data
Exact reasoning traces can teach a student model more than final answers alone. In Anthropic's SWE-bench Verified test, a model trained on full traces scored 80.4%, compared with 71.8% when trained on summaries. The defenses therefore protect not just an answer, but material that can accelerate capability transfer.
If open-weight labs used such data at scale, lower-quality access and more expensive collection will raise their costs. That is the core of Mollick's argument. Public evidence does not reveal how much unauthorized distillation actually contributed to open-weight progress.
A tempting slowdown theory still lacks independent evidence
Both labs measure their defenses on their own systems. OpenAI compares two model and safeguard generations at once, while Anthropic says its new layer has not yet faced a highly sophisticated and persistent attempt. The numbers demonstrate stronger resistance in company tests, not the cause of an entire market's trajectory.
Legitimate knowledge distillation must also be separated from covert extraction of another provider's service. The former is a standard technique used with authorization. The dispute concerns industrial campaigns using fraudulent accounts and transferring capabilities without permission.
The next open-weight generation will reveal the cost of closed access
The useful signals will be benchmarks and technical reports for upcoming open-weight releases. If the gap widens specifically on tasks that benefit from long reasoning traces, Mollick's hypothesis will gain weight. If teams replace the lost source with their own synthetic data or open research, any slowdown may be brief.
The effect on ordinary API customers matters too. Anthropic already charges for some blocked requests, and context restrictions forced changes to some integrations. A defense that stops extractors without tripping legitimate developers will matter more than a striking laboratory curve.
Lilith's verdict
The frontier labs have raised the drawbridge. The next open-weight releases will show who was running on their own legs and who was riding on borrowed reasoning.