OpenAI cuts AI model distillation success rates in new updates
New security upgrades in GPT-6.2 Astra significantly hinder attempts by bad actors to copy the internal reasoning of advanced AI models.

The short version
- Success rate for extracting reasoning from previous requests dropped from 54 percent to 0.2 percent in GPT-6.2 Astra.
- OpenAI is specifically protecting 'chain-of-thought' data, which contains intermediate steps not shown in final answers.
- The company detected multiple unauthorized distillation campaigns this year, enforcing actions against thousands of accounts.
- OpenAI is sharing indicators of compromise with other providers to disrupt coordinated extraction across different platforms.
Why it matters
This move makes it much harder for competitors or bad actors to 'clone' the high-level capabilities of frontier models. It protects sensitive information in fields like cybersecurity and biology from being repurposed for harmful use.
The rise of distillation attacks
OpenAI has implemented significant security upgrades to protect its models from a process known as distillation. This machine-learning technique allows one AI model to learn from the data generated by another model. While this is a common practice for legitimate development, it poses a unique risk when done without permission.
When an unauthorized party uses distillation, they can create a model that mimics the advanced capabilities of a frontier system. However, this new model does not inherit the safety protections built into the original system. This creates a loophole where a model might be able to perform dangerous tasks that the original creator intended to restrict.
As reported by The Tribune, OpenAI warned that these risks are particularly high in sensitive sectors. Specifically, the company highlighted concerns regarding cybersecurity and biological research. In these fields, capabilities designed for defensive purposes could be easily repurposed for malicious activities by someone who has successfully distilled the model's logic.
Protecting the chain-of-thought
A primary target for these attacks is what researchers call the 'chain-of-thought.' This refers to the intermediate reasoning steps an AI takes while working toward a final answer. Even if the final response is safe and clean, the internal reasoning steps can contain much more detailed information.
OpenAI found that this internal data is highly valuable for training competing models. An internal study showed that models trained on exact reasoning traces performed better in cybersecurity benchmarks than those trained on simple summaries. This advantage remained even when the models were given more resources to process information during inference.
Because these traces provide a blueprint for how a model thinks, OpenAI is placing a heavy emphasis on keeping them hidden. By shielding the chain-of-thought, the company aims to prevent others from using its internal logic as a shortcut to build powerful, unregulated AI systems. This is a key part of their strategy to maintain the integrity of their frontier models.
Measuring the impact of safeguards
The company conducted internal evaluations using automated red-teaming to measure the effectiveness of its new defenses. These tests involved trying to extract protected reasoning from previous requests. The results showed a sharp decline in the success rate of these attempts in the newer GPT-6.2 Astra model.
Under the original safeguards of the previous GPT-6 Astra, the success rate for these attacks was 54 percent. With the new upgrades in place, that number fell to 0.2 percent. This represents a significant drop in the ability of attackers to successfully extract the model's internal logic.
The security improvements also addressed attacks targeting reasoning generated in current requests. In these instances, the success rate dropped from 22.9 percent to 0.9 percent. These figures suggest that OpenAI's multi-layered defense strategy is significantly narrowing the window for unauthorized capability extraction.
Enforcement and industry cooperation
OpenAI is using a variety of methods to combat these threats, including model-level safeguards and input and output classifiers. They are also implementing API restrictions and actively investigating coordinated extraction campaigns. The company noted that the pace of these unauthorized distillation efforts increased significantly over the summer.
The scale of the enforcement is substantial, with weekly actions ranging from thousands to tens of thousands of accounts. This indicates a high level of organized activity attempting to bypass the company's security measures. OpenAI is treating these campaigns as a serious threat to the safety of the AI ecosystem.
To address the problem on a broader scale, OpenAI is sharing indicators of compromise with other AI model providers. This cooperation helps the industry identify and disrupt similar activities across different platforms. The goal is to create a unified front against actors who try to reproduce advanced capabilities without the necessary safety guardrails.
National security and global risks
The company framed these security measures as a matter of national security. They argued that foreign actors could potentially reproduce advanced capabilities developed by US companies. Once these capabilities are reproduced, they could be deployed by other entities without any of the safety protections the original developers put in place.
OpenAI is calling for stronger coordination among AI developers, deployment partners, and government bodies. They argue that as advanced AI systems become more widely available, protecting them from unauthorized extraction is essential. This is because the tools developed for legitimate use can quickly become weapons if the underlying logic is stolen.
The move toward stricter controls marks a shift in how frontier models are managed. By focusing on the 'how' of AI reasoning rather than just the 'what' of the final output, OpenAI is trying to create a more secure environment. This approach aims to ensure that the benefits of AI can be shared without creating new, unmonitored risks for the public.
What to watch
- How other AI providers respond to OpenAI's sharing of compromise indicators.
- Whether the drop in success rates holds up against more sophisticated, human-led red-teaming.
- New government regulations specifically targeting 'model distillation' as a security risk.
Try it
Look for 'chain-of-thought' reasoning in your own AI interactions to see how models explain their logic before giving a final answer.
Sources and links
- ReportingThe Tribune: the article we sourced
Read the original at The Tribune
First reported by The Tribune. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.