OpenAI Tightens Protection Against Hidden Reasoning Extraction

OpenAI reports a significant reduction in attacks aimed at extracting the internal reasoning of GPT-6.2 Astra and Sol. The issue goes beyond mere confidentiality: it also concerns model security and the race for distillation.

OpenAI claims to have significantly strengthened the defenses of GPT-6.2 Astra and GPT-6.2 Sol against the extraction of their internal reasoning, also known as protected reasoning. In its automated tests, the success rate of attacks dropped from 54.0% to 0.2% for past queries, and from 22.9% to 0.9% for ongoing queries. This protection matters because these intermediate traces can reveal more than the final answer and can be used to copy capabilities without authorization.

Ce que protège OpenAI, exactement

Protected reasoning refers to the calculation steps a model produces to arrive at an answer. These traces are not always visible to the user, but they may contain information absent from the final output. OpenAI explains that attackers specifically seek to recover these elements to train other systems, notably on coding and defensive cybersecurity tasks. Not very fair play, but very profitable for an adversary.

The company says it has deployed multi-level defenses. First, targeted training to teach the model to recognize extraction attempts. Then, input and output filters for messages and tool calls. Finally, API blocks that prevent older models from reading this protected reasoning. The protections apply to GPT-6.2 Astra and GPT-6.2 Sol, and some improvements are also integrated into GPT-6 Astra.

Des attaques plus difficiles à mener à grande échelle

OpenAI does not just harden the model itself. The company also says it carries out blocking actions against coordinated campaigns, with thousands to tens of thousands of accounts banned each week. According to its report, this pace increased significantly during summer 2026. The stated goal is to reduce the amount of recoverable data, increase the cost of campaigns, and shorten reaction times.

Methodologically, the evaluation cited by OpenAI is an "adaptive black-box" test aggregated over six attacking models. The red team adapts their strategy progressively, while only seeing what an external user would see. In this context, the drop from 54.0% to 0.2% for extraction from a previous query, then from 22.9% to 0.9% on the current query, shows that the barrier has clearly been raised.

Pourquoi la distillation inquiète les éditeurs de modèles

The issue goes beyond OpenAI. The publisher links this announcement to a broader problem: distillation, that is, the unauthorized use of a model’s outputs to train another system. OpenAI says it has observed that exact reasoning traces transfer certain cyber capabilities better than simple summaries, even when the original tasks are authorized. The risk for providers is therefore not only losing content; it is also seeing skills leave.

In its internal study, OpenAI compares two variants of the same student model. One is trained with exact traces from the teacher; the other with user-visible outputs, reasoning summaries, and tool interactions. After an initial supervised pass, the model fed with exact traces performs better. After light reinforcement training, it remains ahead at the largest inference budget tested on the cited benchmarks. OpenAI draws a cautious conclusion: protecting exact traces can limit capacity transfer, including on authorized tasks.

Une réponse qui dépasse le seul périmètre d’OpenAI

The publication is part of a broader sequence. OpenAI had already mentioned, on September 30, 2026, a coordinated campaign observed since July 1, 2026. The company then spoke of 16,000 extraction attempts by more than 4,000 users on July 24 and 25, 2026, with a core attributed to individuals associated with Moonshot AI, while specifying it was uncertain whether all operators belonged to a single actor.

In this field, Anthropic is also conducting its own fight against systematic distillation attacks and says it has banned tens of thousands of accounts. The company claims to succeed in 90 to 99% of cases [to be verified according to the source cited in the text]. OpenAI recommends combining query controls, monitoring tool calls, regression testing, and campaign investigations. Defense is no longer played only on the final answer, but on everything the model lets pass between two messages.

Key points

  • 54.0% to 0.2% on past queries, according to OpenAI.
  • 22.9% to 0.9% on ongoing queries.
  • Thousands to tens of thousands of accounts banned each week.
  • GPT-6.2 Astra and GPT-6.2 Sol are the protected models.
  • OpenAI mentions a coordinated campaign observed since July 1, 2026.
  • Anthropic also claims to fight large-scale distillation.

By the numbers

  • 54.0% — attack success rate on GPT-6 Astra before reinforcement.
  • 0.2% — success rate on GPT-6.2 Astra with new defenses.
  • 22.9% — successful extraction on ongoing queries before update.
  • 0.9% — same test after tightening safeguards.
  • 16,000 — attempted queries on July 24 and 25, 2026, according to OpenAI.

To read