SafetyPublished All news from that day
Ongoing case: Cyber-capable AI models Safety finding · 17 stories
Moonshot says cyber-data cuts lowered Kimi K3.1 Pro's own test score to 81.0%
Moonshot says it removed cyber-related data from two training stages of the open-weight Kimi K3.1 Pro, and plans to release the weights by 4 November.
Check our sources · 10 facts from 1 source
Illustration made with AI for ai notis
Key points
- Moonshot says it identified and removed cyber-related data from both training stages for the open-weight release, while keeping the general-purpose coding tasks.
- On its in-house xdev-bench-hard, Moonshot reports 90.5% solved at pass@3 without these measures and 81.0% for the final Kimi K3.1 Pro; the figures are its own.
- Moonshot says full weights will be released by 4 November 2026, with no licence stated, and that cyber capability stays behind proprietary frontier models available for three months.
What happened
Moonshot AI's post introducing Kimi K3.1 Pro, listed on its blog with the date 24 October 2026, has a section on cybersecurity. Moonshot says it found that cyber-related content in its mid-training data and cyber-related tasks in its reinforcement learning mix drove vulnerability-research performance up as post-training scaled.
For the open-weight release it says it identified and removed cyber-related data from both training stages. On xdev-bench-hard, its own benchmark of vulnerability discovery and exploitation tasks, Moonshot reports that a checkpoint without these measures solved 90.5% of problems at pass@3 and the final K3.1 Pro reached 81.0%.
Moonshot says K3.1 Pro's cyber capability remains behind that of frontier proprietary models that have been publicly available for at least three months.
It states two principles for open weights: it aims not to release weights that would extend the absolute frontier of cyber capability, and it intends to progressively raise the cyber capability available in open weights. The post says full model weights will be released by 4 November 2026.
Underlined sentences link to the facts we checked. Dotted words open a short explainer.
Fact 6 of 10 · from Moonshot AI, 24 Oct 2026
"K3.1 Pro's cyber capability remains behind that of frontier proprietary models that have been publicly available for at least three months."
What it means for you
Our viewThe cyber-data removal is Moonshot's own account, measured on Moonshot's own benchmark, and the weights are a plan until they are published. Its stated aim is not to release open weights that would extend the absolute frontier of cyber capability, and it intends to raise the capability available in open weights progressively.
If you plan to self-host Kimi K3.1 Pro, wait for the weights and the licence, then test its cyber behaviour on your own tasks rather than relying on the vendor's benchmark.
It adds no new facts.
Your reaction
We count reactions per story and day, never who reacted. The counts help us choose what goes in the monthly issue. If you are signed in, your own page shows yours too.
We count reactions per story and day, never who reacted. They help choose what goes in the monthly issue.
Check our sources
Every sentence above is checked against this source.
1 Introducing Kimi K3.1 Pro
Open the source archived copy-
Moonshot AI's blog lists "Kimi K3.1 Pro" with the date 2026-10-24, and its post introducing the model has a section headed Cybersecurity.
citeBlog index card "Kimi K3.1 Pro" dated 2026-10-24; post section heading "Cybersecurity".
-
Moonshot says in that post: "we found that cyber-related content was embedded within the general mid-training corpus, and cyber-related tasks were present in the RL environment mix".
cite"we found that cyber-related content was embedded within the general mid-training corpus, and cyber-related tasks were present in the RL environment mix"
-
Moonshot says: "For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness."
cite"For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness."
-
Moonshot describes xdev-bench-hard as "our in-house benchmark of expert-curated, source-code-based vulnerability discovery and exploitation tasks".
cite"xdev-bench-hard, our in-house benchmark of expert-curated, source-code-based vulnerability discovery and exploitation tasks"
-
Moonshot says that on xdev-bench-hard "the checkpoint without these risk-reduction measures" solves 90.5% of problems at pass@3 and "the final K3.1 Pro reaches 81.0%". The figures are Moonshot's own, from its own benchmark.
cite"On xdev-bench-hard, the checkpoint without these risk-reduction measures — no cyber-related data removal at mid-training and no cyber-related RL environment filtering — solves 90.5% of problems at pass@3; the final K3.1 Pro reaches 81.0%."
-
Moonshot says: "K3.1 Pro's cyber capability remains behind that of frontier proprietary models that have been publicly available for at least three months."
cite"K3.1 Pro's cyber capability remains behind that of frontier proprietary models that have been publicly available for at least three months."
-
Moonshot says: "First, we aim not to release open weights that would extend the absolute frontier of cyber capability."
cite"First, we aim not to release open weights that would extend the absolute frontier of cyber capability."
-
Moonshot says: "Second, we intend to progressively raise the cyber capability available in open weights."
cite"Second, we intend to progressively raise the cyber capability available in open weights."
-
Moonshot says: "Full model weights will be released by Nov 4, 2026." The post states no licence.
cite"Full model weights will be released by Nov 4, 2026."
-
Moonshot says: "During development, we observed vulnerability research performance improving steadily as post-training scaled."
cite"During development, we observed vulnerability research performance improving steadily as post-training scaled."
Every sentence above is checked against this source. Each fact has its own address you can cite.
Source 1 · Moonshot AI · 24 Oct 2026
Introducing Kimi K3.1 Pro
Moonshot AI's blog lists "Kimi K3.1 Pro" with the date 2026-10-24, and its post introducing the model has a section headed Cybersecurity.
Blog index card "Kimi K3.1 Pro" dated 2026-10-24; post section heading "Cybersecurity".
Moonshot says in that post: "we found that cyber-related content was embedded within the general mid-training corpus, and cyber-related tasks were present in the RL environment mix".
"we found that cyber-related content was embedded within the general mid-training corpus, and cyber-related tasks were present in the RL environment mix"
Moonshot says: "For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness."
"For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness."
Show all 10 factsShow fewer facts
Moonshot describes xdev-bench-hard as "our in-house benchmark of expert-curated, source-code-based vulnerability discovery and exploitation tasks".
"xdev-bench-hard, our in-house benchmark of expert-curated, source-code-based vulnerability discovery and exploitation tasks"
Moonshot says that on xdev-bench-hard "the checkpoint without these risk-reduction measures" solves 90.5% of problems at pass@3 and "the final K3.1 Pro reaches 81.0%". The figures are Moonshot's own, from its own benchmark.
"On xdev-bench-hard, the checkpoint without these risk-reduction measures — no cyber-related data removal at mid-training and no cyber-related RL environment filtering — solves 90.5% of problems at pass@3; the final K3.1 Pro reaches 81.0%."
Moonshot says: "K3.1 Pro's cyber capability remains behind that of frontier proprietary models that have been publicly available for at least three months."
"K3.1 Pro's cyber capability remains behind that of frontier proprietary models that have been publicly available for at least three months."
Moonshot says: "First, we aim not to release open weights that would extend the absolute frontier of cyber capability."
"First, we aim not to release open weights that would extend the absolute frontier of cyber capability."
Moonshot says: "Second, we intend to progressively raise the cyber capability available in open weights."
"Second, we intend to progressively raise the cyber capability available in open weights."
Moonshot says: "Full model weights will be released by Nov 4, 2026." The post states no licence.
"Full model weights will be released by Nov 4, 2026."
Moonshot says: "During development, we observed vulnerability research performance improving steadily as post-training scaled."
"During development, we observed vulnerability research performance improving steadily as post-training scaled."
Topics
The morning email
On the mornings we publish: the three top stories and up to four short ones. Free.
We email you a link to confirm. An issue may include one sponsor, always labelled Sponsored · Advertisement. Our emails count opens and clicks, not who made them. Unsubscribe in one click. What we keep