Model

OpenAI says attackers copied GPT-6 Astra's hidden reasoning on 54% of tries; its fix cuts that to 0.2%

OpenAI's November 20 disclosure puts a hard number on how copyable a frontier model was, and pairs the fix with a call for federal enforcement. The labs it names, Moonshot AI and DeepSeek among them, are where the money moves.

Vincent JiangVincent Jiangthree-minute read
Share
Close-up portrait of Sam Altman in a dark suit and tie
Sam Altman, chief executive of OpenAI, in New Delhi on February 20, 2026, at the India AI Impact Summit.

The half of the model nobody sees

For most of this year, an attacker working through OpenAI's$1.18T — OpenAI, private, latest valuation $1.18T ordinary API could recover the reasoning GPT-6 Astra never shows anyone: the chain-of-thought steps the model generates on the way to an answer 1. In OpenAI's own red-teaming, where automated attackers adapt to whatever defenses they meet, that extraction worked on 54.0 of 100 attempts against GPT-6 Astra's original safeguards 1. Hidden traces are the training data a rival lab wants most; OpenAI says its weekly enforcement against extraction campaigns has run from thousands to tens of thousands of accounts, rising sharply through the summer 1.

The fix is 0.2%; the ask is Congress

Updated safeguards in GPT-6.2 Astra cut the earlier-request attack to 0.2%, and extraction from the current request fell from 22.9% to 0.9%, averaged across six attacker models 1.

Extraction success fell from 54.0% to 0.2% after OpenAI's safeguard update

New safeguards cut protected-reasoning extraction success below 1% on both attack types

  • GPT-6 Astra, original safeguards
  • GPT-6.2 Astra, updated safeguards
0%20%40%60%Earlier requestCurrent request0.9%
Data
GPT-6 Astra, original safeguardsGPT-6.2 Astra, updated safeguards
Earlier request54%0.2%
Current request22.9%0.9%
Average attack success rate in adaptive black-box red-teaming, aggregated across six attacker models; lower is better.1

The post does not stop at code. OpenAI backs legislation to sanction foreign actors that illicitly extract US model capabilities, names the Chip Security Act, the AI OVERWATCH Act and the MATCH Act, and says distillation defenses belong beside chip export controls 1.

Washington is already in: a September 9 advisory from CISA, the NSA and the FBI, the US cyber and intelligence agencies, says six Chinese firms, including DeepSeek, Moonshot AI and Alibaba−1.82% — Alibaba, down 1.82 percent today, pulled billions of tokens from US frontier models since at least late 2024, and assesses Beijing likely knows 2. OpenAI wants to be paid twice for the same capability: once in tokens, once in law.

The copy campaign has receipts

The September campaign began July 1, peaked at 16,000 extraction requests in two days across more than 4,000 users, and was disrupted by July 28; OpenAI tied a core cluster of it to individuals at Moonshot AI, the lab behind the Kimi models 3.

Reuters, citing The Information, reported that Moonshot's Kimi K3.1 Pro and DeepSeek's V4.1-Flash 1017, launched October 17, benefited from distilling Astra; neither firm has answered the allegation 4. Anthropic$350B — Anthropic, private, latest valuation $350B counted nearly 200 million exchanges across five campaigns, 151 million of them attributed to Alibaba as training material for its Qwen models 5. OpenAI's own study shows why the traces are worth stealing: students fine-tuned on exact reasoning beat students trained on what users actually see 1.

Beijing calls it a smear

China's foreign ministry told Washington to "refrain from making unfounded accusations or smears," and no named firm has publicly answered 34. The loudest objection on Hacker News ran the other way: OpenAI built its business on "the organized illegal downloading of copyrighted texts" and now wants the state to police copying 6. Defenses alone also decay, by OpenAI's own account, since the September campaign changed its methods as it went 3. Anthropic, for its part, says countermeasures cut data harvested in distillation attacks by roughly 90% to 99% without any new law 4.

If open weights stall, distillation was load-bearing

OpenAI says more campaign disclosures are coming shortly 1. The tell is the one Ethan Mollick, the Wharton professor, posted within minutes: if open-weights progress now stalls, distillation was load-bearing all along 7. Alibaba (NYSE: BABA) is the listed proxy for the named camp; OpenAI and Anthropic are private. The next reading is public and dated: whatever Kimi, DeepSeek and Qwen ship next.

Get Model by email, free

The AI labs, their models, and the business behind them. In our daily email, every weekday.

Loading BABA

Reader comments

0 comments

    Model

    Mira Murati's Thinking Machines ships a 26-billion-parameter model that runs offline on a MacBook

    The 26-billion-parameter open-weights agent shipped hours after OpenAI said attackers had extracted GPT-6 Astra's hidden reasoning at a 54% success rate and asked Congress for new enforcement powers. Together the two moves frame the scarcity bet behind frontier token prices and the market's newest trillion-dollar listing.

    Model

    Amazon's $188 Billion Anthropic Stake Swings $1.9 Billion for Every $1 of Anthropic Stock

    Two trading days turned Amazon's largest AI holding into a daily mark-to-market item, and Anthropic's second-session slide shows how fast the gain Amazon already booked can reverse.

    Model

    HeyGen sells Avatar 6 while Tavus's Griffin scores 3.83 against a human's 3.92 and still has no price

    HeyGen put Avatar 6 on general availability on November 17 with a public API and first place in New Measure's blind Avatar Arena; Tavus's Griffin, which Tavus says is the highest score ever recorded for an AI on NVIDIA's face-to-face benchmark, is still a preview with no price. Revenue accrues to HeyGen today because Avatar 6 is the model on sale; Tavus's best model is not.

    Model

    Pika's Enact lets creators direct AI video mid-scene, with answers in under a second

    Enact, posted November 18, lets a creator drag an AI actor mid-scene, talk to the cast by name and redo the camera after locking a performance. Pika open-sourced the model underneath and kept the directing layer closed, a bet that value in AI video is moving from model quality to workflow.

    Model

    Anthropic falls 11% toward its $83 IPO price as Microsoft undercuts Claude

    Anthropic fell about 11% Friday morning, back toward the $83 IPO price, as Microsoft's price attack on Claude met the open market. Microsoft holds Anthropic's compute contract, its shares and a cheaper rival product; a leveraged income ETF on the two-day-old stock listed the same morning.

    Model

    Ramp's invoices show Anthropic held 58.7% of US business AI spend vs OpenAI's 37.2% before its record IPO

    Ramp's index of 85,000 US business bill-payers is the first independent check on the record $81.3 billion IPO: the enterprise lead over OpenAI is real, and so is the falling price of what Anthropic sells.

    Subscribe to TheInference for free