POLICY · MODELS · #4013
OpenAI tightens protections to stop extraction of internal reasoning from GPT-6.2 Astra and Sol
OpenAI says it has sharply tightened safeguards against extracting internal reasoning from GPT-6.2 Astra and GPT-6.2 Sol: internal adaptive tests showed attack success rates falling from 54.0% to 0.2% for prior requests and from 22.9% to 0.9% for active requests. The company combines targeted training, input/output filters, API blocks that prevent older models reading protected reasoning, and large-scale account bans, and is bringing selected improvements to GPT-6 Astra.
KEY POINTS
- OpenAI says it has sharply tightened safeguards against extracting internal reasoning from GPT-6.2 Astra and GPT-6.2 Sol: internal adaptive tests showed attack success rates falling from 54.0% to 0.2% for prior requests and from 22.9% to 0.9% for active requests.
- The company combines targeted training, input/output filters, API blocks that prevent older models reading protected reasoning, and large-scale account bans, and is bringing selected improvements to GPT-6 Astra.
- Reduces the practical risk that attackers can extract proprietary internal reasoning from advanced models, lowering model-theft and safety risks and demonstrating concrete mitigation techniques.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
OpenAI says it has sharply tightened safeguards on GPT-6.2 Astra and GPT-6.2 Sol that reduce successful extractions of protected internal reasoning; it is also publishing new prompt guidance and API changes for developers.
WHY NOW
OpenAI reports large reductions in attack success rates in internal adaptive tests (from 54.0% to 0.2% for prior requests and 22.9% to 0.9% for active requests on Astra), indicating materially lower short-term risk that attackers can extract proprietary internal reasoning. At the same time OpenAI’s API and prompt guidance change developer-facing behavior (models answer more briefly by default, conversation-history editing can discard earlier reasoning), so developers should review integrations as they move to GPT-6.2.
WHO IS AFFECTED
Directly affected per sources: developers and integrators of GPT-6.2 Astra and GPT-6.2 Sol (prompt templates, API behaviors), users whose workflows rely on conversation history or on model output style, and parties attempting to extract internal reasoning (attacker campaigns that OpenAI says it bans accounts to curb).
CONFIRMED
- OpenAI reports it tightened safeguards on GPT-6.2 Astra and GPT-6.2 Sol. - Internal automated adaptive tests reported attack success rates dropping on GPT-6.2 Astra: prior-request attacks from 54.0% to 0.2%; active-request attacks from 22.9% to 0.9%. - Defenses described by OpenAI include targeted training, input and output filters for messages and tool calls, API blocks that prevent older models reading protected reasoning, and large-scale account bans (thousands to tens of thousands weekly). - OpenAI is bringing selected improvements to GPT-6 Astra. - OpenAI published new prompting tips and copy-ready prompt templates for GPT-6.2 Astra and Sol to address briefer answers, unnecessary clarification questions, coding practices, and to instruct final answers to be self-contained; it also changed API behaviors (editing conversation history can discard earlier reasoning; some parameters like "temperature" and "top_p" are removed).
UNCERTAIN
Missing or unclear in the supplied excerpts: - Independent, third-party validation of OpenAI’s internal adaptive-test results (no external test data provided). (missing) - Exact technical details, scope, and permanence of the targeted training, filters, and API blocks (described at a high level only). (missing) - Which specific improvements are being ported to GPT-6 Astra and on what timeline. (missing) - How durable these defenses are against future or novel distillation/distillation-style attacks and coordinated adversary adaptations. (missing) - Precise operational impact on token costs and response quality across real-world apps beyond the stated effect that editing history can discard reasoning. (partial/conflicting: reported effect but magnitude unclear) - Exact counts and criteria for the account bans beyond the summary phrase "thousands to tens of thousands weekly." (partial) These uncertainties are explicitly not resolved in the provided excerpts.
WHAT TO WATCH
Concrete observable signals to watch next (explicitly supported by excerpts): - Independent red-teaming or third-party tests reporting attack success rates against GPT-6.2 Astra and Sol (to validate OpenAI’s internal ASR numbers). (related to source 4757) - OpenAI release notes or documentation showing which specific improvements were ported to GPT-6 Astra and details on API/blocking behavior. - Uptake or discussion in developer documentation/forums about adopting the new prompt templates and their effects on output brevity, self-contained final answers, and coding templates. - API changelog entries and technical docs confirming conversation-history editing semantics, removal of parameters like "temperature" and "top_p", and any stated impacts on token accounting or response quality. - Reporting or disclosure from OpenAI on account-ban volumes and enforcement criteria (to corroborate the stated scale of bans). - External industry reports (e.g., other providers’ statements like Anthropic’s actions) on coordinated responses to distillation attacks for comparative context. Watch for those signals in official docs, third-party audits, and developer community threads.
WHY IT MATTERS
Reduces the practical risk that attackers can extract proprietary internal reasoning from advanced models, lowering model-theft and safety risks and demonstrating concrete mitigation techniques.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
OpenAI says it has sharply tightened safeguards to prevent extraction of internal reasoning from GPT-6.2 Astra and GPT-6.2 Sol.
SUPPORTEDChecked 2026-11-21 · 1 supporting
In internal adaptive/automated tests OpenAI reports attack-success rates dropped for prior requests from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra, and for active requests from 22.9% to 0.9%.
SUPPORTEDChecked 2026-11-21 · 1 supporting
OpenAI describes the defenses as combining targeted training, input/output filters for messages and tool calls, API blocks that stop older models reading protected reasoning, and large-scale account bans.
SUPPORTEDChecked 2026-11-21 · 1 supporting
OpenAI says it is bringing selected improvements from GPT-6.2 to GPT-6 Astra.
SUPPORTEDChecked 2026-11-21 · 1 supporting
SOURCES & TIMELINE
1OpenAI has sharply tightened safeguards against extracting internal reasoning from GPT-6.2 Astra and GPT-6.2 Sol. In internal automated tests, attacks pulling protected reasoning from prior requests dropped from a 54.0 percent success rate on GPT-6 Astra to 0.2 percent on GPT-6.2 Astra. Attacks on active requests fell from 22.9 to 0.9 percent, OpenAI says . The company is now bringing selected improvements to GPT-6 A…
OpenAI advises developers to update GPT-6.2 Astra and Sol system prompts. The models answer more briefly by default and should make assumptions instead of interrupting workflows with questions. For coding, the template instructs the models to reuse existing code, avoid unnecessary abstractions, and write new tests only on request or for specific errors. In the API, editing conversation history discards earlier reas…