What changes in the new defenses
OpenAI has announced that it has sharply strengthened protections against attempts to extract the internal reasoning of its GPT-6.2 Astra and GPT-6.2 Sol models. According to The Decoder, in automated internal tests the success rate of attacks that recovered reasoning protected from previous requests fell from 54.0 percent on GPT-6 Astra to 0.2 percent on GPT-6.2 Astra. For attacks conducted on active requests, the reduction cited is from 22.9 percent to 0.9 percent. OpenAI also specified, according to the source, that some of the selected improvements will be extended to GPT-6 Astra as well.
Two levels of defense and countering coordinated campaigns
The countermeasures described operate on two distinct levels. At the level of the individual request, the systems combine targeted training, filters on inputs and outputs for messages and tool calls, and API blocks that prevent older models from reading protected reasoning. At the level of coordinated campaigns, OpenAI states that it suspends between thousands and tens of thousands of accounts every week.
The significance of these measures lies in the fact that a model's internal reasoning can be a competitive asset and, potentially, a vehicle for reconstructing capabilities that have not been made public. However, the figures cited come from the company's own internal tests: they are therefore self-interested statements, not independent verifications, and should be read as such.
A shared front with Anthropic and the open questions
The phenomenon does not concern OpenAI alone. According to The Decoder, Anthropic is also countering systematic distillation attacks, suspending tens of thousands of accounts to limit the theft of internal reasoning, and has reportedly said it succeeded in 90 to 99 percent of cases. Anthropic has also reportedly called for a coordinated response from industry and policymakers.
The comparison between the two companies shows an approach that is in some respects similar, based on mass account suspensions, but the material does not allow direct comparison of metrics and methodologies: OpenAI publishes attack success rates in adaptive red-teaming tests, while for Anthropic the source refers to a percentage of cases in which the countermeasures reportedly succeeded, without detailing the calculation basis.
Several aspects therefore remain to be verified: the actual robustness of the defenses against external attackers and not only in the company's tests, the criteria by which accounts are identified and suspended, the risk of erroneous suspensions, and the outcome of requests for coordination among sectors and institutions. On these points, the available material offers no further elements.
