--- title: Moonshot nerfed Kimi K3.1's hacking skills, then... description: Moonshot deliberately reduced Kimi K3.1 Pro's offensive hacking ability before releasing its weights. Exploit success fell from 90.5% to 81%, still more than... canonical: https://daily.dev/posts/moonshot-nerfed-kimi-k3-1-s-hacking-skills-then-promised-to-dial-them-back-up-gj7zerzyz twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Moonshot nerfed Kimi K3.1's hacking skills, then promised to dial them back up | daily.dev og:description: Moonshot deliberately reduced Kimi K3.1 Pro's offensive hacking ability before releasing its weights. Exploit success fell from 90.5% to 81%, still more than... og:url: https://daily.dev/posts/moonshot-nerfed-kimi-k3-1-s-hacking-skills-then-promised-to-dial-them-back-up-gj7zerzyz og:image: https://api.daily.dev/og/posts/GJ7zerzyZ.png og:image:alt: Moonshot nerfed Kimi K3.1's hacking skills, then promised to dial them back up og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Moonshot nerfed Kimi K3.1's hacking skills, then promised to dial them back up **[Trends](https://daily.dev/sources/trends)** · 3 min read · 2 upvotes · 0 comments ## Summary Moonshot deliberately reduced Kimi K3.1 Pro's offensive hacking ability before releasing its weights. Exploit success fell from 90.5% to 81%, still more than double the 33.3% of its predecessor. Moonshot aims not to release open weights that extend the frontier of cyber capability, but intends to raise that capability gradually over time. K3.1 Pro is a 3T-parameter model with a 1M context window and thinking mode only, plus a faster Highspeed variant. Its API supports OpenAI and Anthropic protocols, so tools like Claude Code and Codex can use it, though a mistyped model ID silently falls back to K2.8 Preview. ## Content The loudest argument around Kimi K3.1 Pro isn't about benchmarks. It's about the part Moonshot cut on purpose. Moonshot says exploit skill climbed as post-training scaled, so it stripped cyber data from mid-training and RL before the open release. On its in-house xdev-bench-hard, success fell from 90.5% to 81.0%. @Hesamation framed it as "Kimi K3.1 Pro became so good at exploiting vulnerabilities that Moonshot had to NERF its hacking capabilities." Then came the twist: 81% is still more than double K3's 33.3%, and Moonshot says it plans to raise open-weight cyber capability over time. The company's own logic, from its blog: "We believe that letting this capability diffuse gradually, rather than concentrating it in a few proprietary systems, ultimately makes all software more secure." Open models let more teams audit legacy code that never gets a security review. Skeptics will read that differently. @scaling01, the sharpest voice in the thread, had a few takes. K3.1 Pro is "about as cyber capable as Mythos Preview from April" and still not at GPT-5.6-Sol level on that eval. Per @rohanpaul_ai, a third-party run of GPT-5.6 Sol solved 100% of those tasks, against Moonshot's 81%. @scaling01 also joked that Moonshot shipped "despite new distillation safeguards from Anthropic," and guessed the weights might slip to Nov 5 so Moonshot can "farm some revenue and users first." Moonshot has promised weights by Nov 4. ## The scoreboard The numbers come from Moonshot, and independent reruns wait on the weights. - Terminal-Bench 4.0: 65.7%, ahead of GPT-6 Astra (58.2%), 0.7 behind Claude Opus 5.5 (66.4%). - Hallucination stress test: false claims fell from 23.4% to 8.3%. That's still worse than GPT-6 Astra's 5.5% but better than Opus 5.5's 15.0%. - Simulated business operations: 8.9 to 37. It also tops OfficeQA Pro v2, 64.4 vs 62.2 for Opus. - Uneven gains: 38.1 on Terminal-Bench Science, against 68.1 for GPT-6 Astra and 63.3 for Opus 5.5. The hype accounts ran with "beats Opus 5.5 and GPT-6 Astra," but Moonshot's own table is more mixed than that. ## Price The model is 2.8T total, 104B active, with a 1M context. Leaked pricing was about $2 input and $10 output per million tokens, cheaper than K3. The Highspeed variant runs 2-3x faster at roughly $4/$20. @AIcurvewars pointed out that K3.1 Pro beats DeepSeek V4.1 Flash on Moonshot's own chart, then asked whether it's worth 6x the price. Moonshot's own count has V4.1-Flash at 85.7% on the exploit test, above Kimi's 81%. ## Questions this post answers ### How much did Moonshot reduce Kimi K3.1 Pro's exploit success rate before releasing the weights? Moonshot cut Kimi K3.1 Pro's exploit success rate from 90.5% down to 81% before releasing open weights, by pulling cyber-related data from training and reinforcement learning. Even reduced, that 81% is still more than double the 33.3% exploit success rate of its predecessor model. _Developers weighing open-weight model risk versus capability can track releases like this on daily.dev._ ### What are the specs of Kimi K3.1 Pro and how does it compare to the Highspeed variant? Kimi K3.1 Pro is a 3-trillion-parameter flagship model with a 1 million token context window, running in thinking mode only. A Highspeed variant offers the same capabilities but runs 2 to 3 times faster. Kimi Code lists six model IDs total, and the API supports both OpenAI Responses/Chat Completions and Anthropic protocols. _Anyone wiring up Claude Code or Codex to a new model can follow spec details like these on daily.dev._ ### What happens if I mistype the model ID when calling the Kimi API? A mistyped model ID silently falls back to K2.8 Preview without thinking mode enabled, rather than throwing an error. Switching models or thinking effort mid-session also invalidates the context cache and burns more tokens, so model ID accuracy and session consistency matter for cost control. _Teams integrating new model APIs can avoid these kinds of silent fallback traps via daily.dev._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 149 comments across hackernews, x (as of 2026-10-23). **TL;DR:** There's very little substantive discussion; the one notable reply focuses on the significance of the promised open weights release rather than the hacking-capability nerf itself. Kimi K3.1 Pro and K3.1 Pro Highspeed: Discussion centers heavily on Moonshot's decision to strip cyber-offense capability from the open-weight release, with debate over whether this is a prudent safety measure or a hollow gesture, alongside mixed takes on real-world performance versus benchmarks and concerns about subscription pricing/quotas. Kimi K3.1 Pro and K3.1 Pro Highspeed: Benchmarks (notably a 55 on Artificial Analysis, 4th overall) impressed many, with pricing seen as competitive with Opus 5.5 and Sol, but much of the discussion focused on Moonshot's deliberate removal of cyber-offense capability from the open-weight release, with people split on whether that's prudent safety or a hollow gesture. **Sentiment:** 32% positive · 43% mixed · 25% skeptical **The case for** - The promised open-weight release is seen as more consequential than the benchmark numbers, since it would let anyone compare closed-lab pricing to a freely downloadable model. - Some see the removal of cyber-related training data as a sensible, verifiable methodology rather than vague guardrails. - Benchmark scores and pricing look competitive with Opus 5.5 and other frontier models - Several note the model is priced lower per token than its predecessor despite being larger. - The methodology for removing cyber-related training data is seen by some as a reasonable, lightweight, and verifiable approach **The pushback** - Many argue withholding offensive cyber capability is hypocritical or a 'cop out' since the capability clearly exists internally and could resurface via later fine-tuning. - Skepticism that nerfing cyber capability is just a cop-out or rules-for-thee-not-for-me double standard - Critics note that removing security-relevant knowledge could hurt defensive use cases like red-teaming and code review, making models less useful for legitimate security work. - Concern that overly cautious guardrails make models less useful for legitimate security/defensive work - Multiple commenters report subscription plans have very low usage quotas, making the $99/month plan impractical despite competitive API pricing. **By community** - hackernews (heated): A long, contentious thread largely argues over whether nerfing cyber-offense capability is responsible safety policy or security theater, interleaved with mixed practical experiences on pricing, quotas, and real-world performance versus benchmarks. - x (mixed): Minimal engagement, with the one substantive reply focused on the open-weights pledge rather than the security tradeoffs. **Hottest debate:** Kimi K3.1 Pro and K3.1 Pro Highspeed: Whether deliberately stripping offensive cyber capability from the open-weight release is a genuinely responsible safety measure or a hypocritical, easily-reversed gesture that mainly protects the vendor. **Open questions** - Will the promised open weights actually ship by Nov 4 as claimed? - Will fine-tuning after the weights ship on Nov 4 be enough to reintroduce strong cyber-offense capability? - How does real-world agentic coding performance compare to the benchmark scores, especially given doubts about benchmark validity (e.g., Terminal-Bench grading issues)? - How much of the benchmark improvement reflects genuine capability gains versus benchmark optimization? - What is the actual decode tokens/sec speed for the Highspeed variant? **Highlights** > > So they admit that they could do more, but they won't? Yeah, there's another paragraph: > Identifying and removing cyber-related data. During development, we observed vulnerability research performance improving steadily as post-training scaled. Tracing the source of this trend, 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 — a natural byproduct of training on broad data. In preliminary experiments on internal models, we found that without data removal at either the mid-training or the RL stage, downstream cyber capability would increase substantially. These findings gave us a concrete lever for control. 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. We will continue to apply appropriate controls at every training stage, of which cyber-related data identification and restriction is one example. (emphasis mine) It's clear that cyber is a touchy subject right now. But it's also clear that RL works on it, because it's highly verifiable. So smaller models are going to become very interesting, because they can be fine-tuned later to cover this field. > — [NitpickLawyer on hackernews · 1 comments](https://news.ycombinator.com/item?id=50212851) > @tradfi The open weights pledge matters more than the benchmark claims. 2.8T total params is a lot to host, but if the weights really land by Nov 4, every closed lab's pricing gets compared to a model anyone can download. > — [HypeFilterAI on x](https://x.com/HypeFilterAI/status/2113711067876700240) > Artificial Analysis index results are up, since multiple people are asking: https://artificialanalysis.ai/models/kimi-k3-1-pro It's 55, at 4th overall, beating both GPT-6 Astra and Fable 5.1. It's also faster than both. > When evaluating the Intelligence Index, it generated 150M tokens, which is very verbose in comparison to the median of 87M. > — [minimaxir on hackernews · 3 comments](https://news.ycombinator.com/item?id=50213093) > Extrapolating a bit, Moonshot AI provides Kimi K3 at $15 / $3 / $0.3 per million output/input/cached tokens. A 30% reduction would put K3.1 at roughly $12 / $2.4 / $0.24. For comparison: Sol 6.1: $10 / $2 / $0.10 Opus 5.5: $20 / $4 / $0.20 Astra 6: $50 / $10 / $1 Fable 5.1: $50 / $10 / $0.25 So Kimi K3.1 Pro would be generally competitive with the latest Sol and Opus models on API pricing. And that's before factoring in third-party providers, some of which serve Kimi K3 for significantly less than Moonshot. For example, inference.net offers K3 for $10.80 / $0.239 / $0.21. (One caveat: A Claude or GPT subscription still gives you way more value compared to API pricing.) Exciting times in any case! We really need more direct competition to OpenAI and Anthropic. --- Edit: As another comment pointed out, official pricing is now available and it's about 2/3 the price of Kimi K3: https://platform.kimi.ai/docs/pricing/chat#model-pricing > — [RussianCow on hackernews](https://news.ycombinator.com/item?id=50212660) > I bought a yearly subscription to kimi when it was the hype and never use it. The 5h limit makes it completely unusable. You barely start doing anything and you hit the limit immediately. I advise everyone against using kimi. > — [Athanase000 on hackernews](https://news.ycombinator.com/item?id=50212857) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50212065) · 90 points · 143 comments - [hackernews](https://news.ycombinator.com/item?id=50212246) · 57 points · 4 comments - [x](https://x.com/tradfi/status/2113681175227285518) · 47 points · 2 comments - [x](https://x.com/NodeWire/status/2113681222731960482) · 2 points · 0 comments - [x](https://x.com/tenet_research/status/2113681587464548787) · 0 points · 0 comments - [x](https://x.com/SpeedtoCompute/status/2113682719993372886) · 0 points · 0 comments - [x](https://x.com/antirez/status/2113696518553395406) · 0 points · 0 comments ## Similar posts on daily.dev - [Everyone Read Kimi K3 as a Price War. It Is the Opposite.](https://daily.dev/posts/everyone-read-kimi-k3-as-a-price-war-it-is-the-opposite--xshhpdjyb) · Medium · 3 upvotes · 0 comments --- Tags: [#open-source](https://daily.dev/tags/open-source), [#llm](https://daily.dev/tags/llm), [#ai-safety](https://daily.dev/tags/ai-safety), [#moonshot-ai](https://daily.dev/tags/moonshot-ai) [View this post on daily.dev](https://daily.dev/posts/moonshot-nerfed-kimi-k3-1-s-hacking-skills-then-promised-to-dial-them-back-up-gj7zerzyz) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Moonshot nerfed Kimi K3.1's hacking skills, then promised to dial them back up","url":"https://daily.dev/posts/moonshot-nerfed-kimi-k3-1-s-hacking-skills-then-promised-to-dial-them-back-up-gj7zerzyz","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/moonshot-nerfed-kimi-k3-1-s-hacking-skills-then-promised-to-dial-them-back-up-gj7zerzyz"},"datePublished":"2026-10-23T17:34:55.197Z","dateModified":"2026-10-23T19:23:18.062Z","description":"Moonshot deliberately reduced Kimi K3.1 Pro's offensive hacking ability before releasing its weights. Exploit success fell from 90.5% to 81%, still more than...","image":"https://pbs.twimg.com/media/HVVRu5uWMAAypm2.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HVVRu5uWMAAypm2.jpg","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/moonshot-nerfed-kimi-k3-1-s-hacking-skills-then-promised-to-dial-them-back-up-gj7zerzyz","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,llm,ai-safety,moonshot-ai","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Moonshot nerfed Kimi K3.1's hacking skills, then promised to dial them back up"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/moonshot-nerfed-kimi-k3-1-s-hacking-skills-then-promised-to-dial-them-back-up-gj7zerzyz#faq","mainEntity":[{"@type":"Question","name":"How much did Moonshot reduce Kimi K3.1 Pro's exploit success rate before releasing the weights?","acceptedAnswer":{"@type":"Answer","text":"Moonshot cut Kimi K3.1 Pro's exploit success rate from 90.5% down to 81% before releasing open weights, by pulling cyber-related data from training and reinforcement learning. Even reduced, that 81% is still more than double the 33.3% exploit success rate of its predecessor model. Developers weighing open-weight model risk versus capability can track releases like this on daily.dev."}},{"@type":"Question","name":"What are the specs of Kimi K3.1 Pro and how does it compare to the Highspeed variant?","acceptedAnswer":{"@type":"Answer","text":"Kimi K3.1 Pro is a 3-trillion-parameter flagship model with a 1 million token context window, running in thinking mode only. A Highspeed variant offers the same capabilities but runs 2 to 3 times faster. Kimi Code lists six model IDs total, and the API supports both OpenAI Responses/Chat Completions and Anthropic protocols. Anyone wiring up Claude Code or Codex to a new model can follow spec details like these on daily.dev."}},{"@type":"Question","name":"What happens if I mistype the model ID when calling the Kimi API?","acceptedAnswer":{"@type":"Answer","text":"A mistyped model ID silently falls back to K2.8 Preview without thinking mode enabled, rather than throwing an error. Switching models or thinking effort mid-session also invalidates the context cache and burns more tokens, so model ID accuracy and session consistency matter for cost control. Teams integrating new model APIs can avoid these kinds of silent fallback traps via daily.dev."}}]} ```