Attack Rates Fall as OpenAI Limits AI Reasoning Extraction

OpenAI says safeguards in GPT-6.2 Astra and GPT-6.2 Sol sharply reduced attempts to extract protected internal reasoning in automated tests. The company combines request-level defenses with account bans against coordinated campaigns, and is bringing selected improvements to GPT-6 Astra.

WTF Index TERMINATOR
◄ Terminator 2 Idiocracy 0 ►

The story describes efforts to limit extraction of protected AI reasoning and coordinated misuse, a mild lean toward controlling AI risks.

Attack Rates Fall as OpenAI Limits AI Reasoning Extraction

OpenAI says it has made it substantially harder to extract protected internal reasoning from its latest AI models. In the company’s internal automated tests, attack success rates against GPT-6.2 Astra fell to a small fraction of the rates recorded for GPT-6 Astra.

Tests show a sharp drop in successful attacks

For attacks seeking reasoning from prior requests, the reported success rate fell from 54.0 percent on GPT-6 Astra to 0.2 percent on GPT-6.2 Astra. For attempts targeting active requests, the rate declined from 22.9 to 0.9 percent.

The results are from adaptive black-box red teaming aggregated across six attacker models. The reported measure is the attack success rate, or ASR; lower values mean fewer attempts succeeded. These figures describe OpenAI’s internal tests, rather than a guarantee that every effort to obtain protected reasoning will fail.

Defenses operate within requests and across campaigns

OpenAI describes a layered approach. For individual requests, its systems use targeted training alongside input and output filters. Those filters apply to messages and tool calls, while API blocks prevent older models from reading protected reasoning.

This combination addresses different points where an extraction attempt could occur. Training is intended to shape model behavior, filters screen what enters and leaves a request, and API restrictions limit access between models. The source does not provide separate effectiveness figures for each of these measures.

OpenAI also says it acts against organized activity beyond a single request. The company bans thousands to tens of thousands of accounts weekly in response to coordinated campaigns. That account-level response complements the technical controls applied to model interactions.

Selected protections are coming to GPT-6 Astra

The company says selected improvements are also being brought to GPT-6 Astra. That extends at least some of the newer safeguards beyond GPT-6.2 Astra and GPT-6.2 Sol, though the source does not specify which improvements will carry over or when they will be available.

The distinction matters because the reported test results concern GPT-6 Astra and GPT-6.2 Astra, while the company’s statement about selected improvements names GPT-6 Astra. The figures should not be read as a published test result for every model or every protection.

Other labs are confronting similar extraction efforts

The article also reports that Anthropic is banning tens of thousands of accounts to counter systematic distillation attacks aimed at obtaining internal reasoning. It says those efforts are reportedly succeeding in 90 to 99 percent of cases, and that Anthropic has called for a coordinated response from industry and policymakers.

Taken together, the measures described point to two parts of the problem: reducing the success of individual extraction attempts and responding to campaigns that use many accounts. OpenAI’s reported reductions show how much its own test results changed between GPT-6 Astra and GPT-6.2 Astra, while its weekly bans address activity at a broader scale.