Brief IA
›

OpenAI Strengthens Protection for GPT-6.2's Reasoning

Models & LLM·Tom Levy·

OpenAI Strengthens Protection for GPT-6.2's Reasoning

OpenAI Strengthens Protection for GPT-6.2's Reasoning
Key Takeaways
1OpenAI strengthens the protections of GPT-6.2 Astra and GPT-6.2 Sol against the extraction of their internal reasoning.
2In its internal tests, attacks targeting a previous request succeed at 0.2% on GPT-6.2 Astra, compared to 54.0% on GPT-6 Astra.
3OpenAI bans thousands to tens of thousands of accounts each week in response to coordinated campaigns.
Why it matters : these defenses aim to prevent access to the protected internal reasoning of the models.

OpenAI strengthens the defenses of GPT-6.2 Astra and GPT-6.2 Sol against the extraction of their internal reasoning. In its internal tests, the success rate of attacks targeting a previous request drops from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra. Some improvements also need to be applied to GPT-6 Astra.

OpenAI and Anthropic ban accounts

In response to coordinated campaigns, OpenAI bans thousands to tens of thousands of accounts each week. Anthropic is also combating systematic distillation and has banned tens of thousands of accounts to limit the theft of internal reasoning.

These measures at Anthropic seem to succeed in 90 to 99% of cases. The company has also called on industry leaders and policymakers to organize a joint response.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Filters and tests to protect requests

For each request, OpenAI combines targeted training and filters applied to messages and tool calls, both in input and output. Restrictions in the API also prevent older models from accessing protected reasoning.

The company compares GPT-6 Astra, equipped with initial defenses, to GPT-6.2 Astra, which has the new protections, in internal automated tests. For attacks targeting an active request, it reports a success rate of 22.9% on the first model, compared to 0.9% on the second. The results aggregate six attacking models in adaptive black-box tests. The indicator used, ASR, measures the success of attacks: the lower it is, the better the protection.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.