The Circuitry
THE CIRCUITRYYour one-stop source for all tech news
  • Today
  • Feed
  • Events
  • Saved
HOMETODAYFEEDEVENTS
BOOKMARKS
RSS
© 2026 The Circuitry
About UsSourcesContactCorrectionsPrivacy
  • Today
  • Feed
  • Events
  • Saved
  • AAPL
  • MSFT
  • GOOGL
  • META
  • NVDA
  • TSLA
  • AMZN
  • INTC
  • SONY
  • TTWO
  • COIN
  • BTC
  • ETH
  • SOL
  • AAPL
  • MSFT
  • GOOGL
  • META
  • NVDA
  • TSLA
  • AMZN
  • INTC
  • SONY
  • TTWO
  • COIN
  • BTC
  • ETH
  • SOL

Feed

Ad
From The CircuitryVerified tech news, cross-checked.How we verify →
  • OpenAI introduces MedChemBench, with GPT-6.2 Astra at 84.5%Tech
  • Cairn agents found a zero-day and took control of a serverTech
  • Google finds an AI agent that built exploits with no human in the loopTech
  • Suspected China-linked group exploited vCenter before Nov. 19Security
  • TikTok launches an AI checker and starts a global likeness rolloutTech
  • Falcon 9 launches USSF-317 and lands the boosterTech
  • Anthropic mathematician says he believes he has a Riemann proofTech
  • Google will retire the Nest app and Assistant on home devices in 2027Tech
  • Niron Magnetics will build a $1.8 billion plant in DothanTech
  • Google's Gemini tool flags wildfire dispatches in under 30 secondsTech
Scroll for more
Verification
VERIFIEDConfidence: HIGH
Source identified
Claims cross-referenced
No discrepancies found
Sourcing
1source
OpenAI · track record
18Stories
100%Verified
1630d
All sources →
Ad
From The CircuitryWhy The Circuitry

Verified tech news, cross-checked.

Every story is checked against independent sources before it posts — no rumors dressed up as fact.

How we verify →
From The CircuitryWhy The Circuitry

Verified tech news, cross-checked.

Every story is checked against independent sources before it posts — no rumors dressed up as fact.

How we verify →
Home/Tech/OpenAI cuts reasoning-extraction success to 0.2% on GPT-6.2 Astra
VERIFIEDBy Xavier Rivera·Friday, November 20, 2026 at 4:11 PM AST·2.5 min read

OpenAI cuts reasoning-extraction success to 0.2% on GPT-6.2 Astra

OpenAI has cut the average success rate of attacks that extract protected reasoning from an earlier request to 0.2% on GPT-6.2 Astra, from 54.0% on GPT-6 Astra. Current-request attacks are at 0.9%, and selected safeguard improvements are being applied back to GPT-6 Astra.

Source:OpenAI
Post
OpenAI cuts reasoning-extraction success to 0.2% on GPT-6.2 Astra
Ad
From The CircuitryVerified tech news, cross-checked.How we verify →
Ad
From The CircuitryWhy The Circuitry

Verified tech news, cross-checked.

Every story is checked against independent sources before it posts — no rumors dressed up as fact.

How we verify →
TL;DRAI · 60 sec read

OpenAI has cut the success rate of attacks that extract protected reasoning from an earlier request to 0.2% on GPT-6.2 Astra, down from 54.0% on GPT-6 Astra. Current-request attacks are at 0.9%. The update also covers GPT-6.2 Sol, binds hidden reasoning to one conversation, and is being applied back to GPT-6 Astra. Weekly enforcement against distillation campaigns reaches thousands of accounts.

Morning Brief

Get the daily OpenAI brief

Every morning, in your inbox. What's in The Brief →

OpenAI has cut the rate at which testers can pull hidden reasoning out of GPT-6.2 Astra. In an internal adaptive test, the success rate for recovering reasoning from an earlier request is 0.2%, down from 54.0% under the original GPT-6 Astra safeguards.

Current-request attacks are at 0.9%. The same black-box test measures reasoning generated for the request in progress, and that rate is down from 22.9% on GPT-6 Astra. OpenAI says both figures are averages across six attacker models, and that each attack type on the updated Astra model stays under 1%. Selected safeguard improvements are now being applied to GPT-6 Astra as well.
Those traces can hold detail the final reply leaves out, which makes them useful training data for a second model.

The target is reasoning the user never sees. OpenAI treats protected reasoning as the intermediate steps a model generates while working toward an answer. Those traces can hold detail the final reply leaves out, which makes them useful training data for a second model. The new defenses cover GPT-6.2 Astra plus Sol, and they sit closer to the controls already used for biology and cybersecurity.

OpenAI@OpenAI

View on X

Three layers sit on each request. Training teaches the models to spot extraction attempts and refuse to hand over protected reasoning, while still answering ordinary requests for explanations. Input classifiers flag extraction attempts, and output classifiers look for leaks in replies and in tool-call arguments, including streamed replies. A leak in a tool argument can expose reasoning even when the final answer does not.
From The CircuitryThe Feed — live briefs across tech, all day.See what’s happening →

The API now binds reasoning to one conversation. On those two models, earlier reasoning carries forward as new messages arrive. If an earlier message is rewritten, the API drops the reasoning that followed it. Older model families with weaker safeguards cannot read that protected reasoning, and OpenAI has shut a route that let one account surface another user's encrypted reasoning.
If an earlier message is rewritten, the API drops the reasoning that followed it.

Weekly enforcement reaches thousands of accounts. The weekly tally this year has run from the thousands of accounts into the tens of thousands, and that pace is significantly higher after this past summer. OpenAI passes indicators from those investigations to other model providers, which can then look for related abuse on their own services. A September report attributed a core part of one campaign to people linked with Moonshot AI, which develops Kimi.
Ad
From The CircuitryWhy The Circuitry

Verified tech news, cross-checked.

Every story is checked against independent sources before it posts — no rumors dressed up as fact.

How we verify →
Exact reasoning traces transferred more cyber skill. In an internal study, a student trained on the teacher's exact traces generally beat a student trained on user-visible summaries after supervised fine-tuning. After a further round of reinforcement learning, the trace-trained student remained ahead when testers gave it the biggest inference budget in the study. OpenAI plans to share more soon on campaigns it has disrupted.
Why this mattersAI · ~100 words

Tap a lens to see what this story means for you.

Morning Brief

Liked this? The Brief brings you the whole day in tech, verified, every morning.

Two minutes, free forever. What's in The Brief →

Reader-supported
DonateBuy me a coffee →Follow@thecircuitry_ →Follow@thecircuitry.to →
HELP US IMPROVE
Grand Theft Auto VILive hubUS store unlocks Thu Nov 19, 1:00 AM ASTFirst unlock: New Zealand · Wed 7:00 AM ASTLaunch hub →Image: Rockstar Games
From The Circuitry

See what’s happening right now

The Feed runs all day — short, verified briefs the moment they break.

Open the Feed →
From The Circuitry

Follow @thecircuitry_

Every story we publish, as it happens. No noise between.

Follow on X ↗On Bluesky ↗

Reader-supported

The Circuitry is a passion project I've always wanted to build, and I love the work behind it.

Running it costs real money. APIs, hosting, time. To keep improving the site and growing this into something useful for everyone, those costs have to be covered.

Any contribution is appreciated. If not, no pressure. Thanks for reading.

Buy me a coffee
openaiaisecurity
Grand Theft Auto VILive hubUS store unlocks Thu Nov 19, 1:00 AM ASTFirst unlock: New Zealand · Wed 7:00 AM ASTLaunch hub →Image: Rockstar Games
More inTech
  • Falcon 9 launches USSF-317 and lands the booster

    Tech · 8h
  • Anthropic mathematician says he believes he has a Riemann proof

    Tech · 21h
  • Google will retire the Nest app and Assistant on home devices in 2027

    Tech · 23h
SupportThe Work

The Circuitry is reader-supported. If you find the daily brief useful, you can buy me a coffee to keep it going.

Buy a coffee →
Ad
From The CircuitryWhy The Circuitry

Verified tech news, cross-checked.

Every story is checked against independent sources before it posts — no rumors dressed up as fact.

How we verify →
From The CircuitryWhy The Circuitry

Verified tech news, cross-checked.

Every story is checked against independent sources before it posts — no rumors dressed up as fact.

How we verify →

MORE IN THIS BEAT

All Tech →
  • Tech·Nov 20

    OpenAI introduces MedChemBench, with GPT-6.2 Astra at 84.5%

    OpenAI has introduced MedChemBench, a 99-problem test of how models read molecules, explain lab results, and choose the next compound. At the xhigh setting, GPT-6.2 Astra scored 84.5%, while compound choice and property forecasts remain much weaker than structure work.

  • Tech·Nov 20

    Cairn agents found a zero-day and took control of a server

    Autonomous Cairn agents found an unpatched flaw in a medical imaging app and took control of a server in a university campaign that began September 8, 2026. Amazon rates a Chinese-nexus link as moderate confidence, and the project advisory still lists no patched version.

  • Tech·Nov 16

    Google finds an AI agent that built exploits with no human in the loop

    Google has identified a framework that drafted and ran exploits with no person in the loop. The same report says zero-day exploits rose to 22 in August, and stolen model access reached $1 million in the worst cases.

  • Tech·Nov 20

    Suspected China-linked group exploited vCenter before Nov. 19

    A suspected China-linked group exploited VMware vCenter bug CVE-2026-123890 in early September, before the flaw was disclosed on Nov. 19. Federal agencies have until Nov. 23 to apply the vendor's mitigations for the known-exploited flaw.

  • Tech·Nov 20

    TikTok launches an AI checker and starts a global likeness rollout

    TikTok has launched a public AI checker for images and videos made with its own tools and has started a global opt-in likeness rollout. The checker is live without an account, while the likeness tool is limited to eligible creators and a notice does not by itself remove a post.