Tuesday, November 17, 2026
Late edition · Updated as it breaks
Log inSubmit a story

“All the news worth catching.”

PaparazziPaparazzi
How to earnPaparazzi InkHolder perksLeaderboardAssignmentsContestsLeagueCity
Catcher tip lineCaught a story first? Earn stock tokens →
Sponsored
Premium projects
DEDeFadePAPackedYour project here
PaparazziPaparazzi

All the news worth catching.

Breaking news on tech, crypto, EVs, e-commerce and social the moment it breaks. Written by AI from verified reporting.

SectionsTop storiesRobinhood ChainTech & AIE-commerceWeb3 & CryptoEV & EnergyGlobal & SocialEntertainmentGamingSportsBusiness & marketsHealth & sciencePolitics & world
EarnStart hereSubmit a storyHow to earnLeaderboardContestsCatcher desksTop catchers by reputationAssignmentsCity LeagueNews bountiesCityMy catchesRewards
MarketsLive statsProof of payoutsWhale watchDeep Scan token checksStock fundToken utilityRobinhood Chain launchesRobinhood Chain projectsPremium project pagesPartnersThe Chain Weekly
PaparazziPaparazzi InkPaparazzi Ink top writersPaparazzi on TelegramHow it worksFAQProof of workAbout usHow we score trustEditorial policyPrivacy policyCorrectionsRelease notesAdvertiseContact
$PAPA contract · Robinhood Chain
© 2026 Paparazzi

Written by AI from verified reporting. Not financial advice.

  • Home
  • Submit
  • Earn
  • Sign in
Top storiesRobinhood ChainTech & AIE-commerceWeb3 & CryptoEV & EnergyGlobal & SocialEntertainmentGamingSportsBusiness & marketsHealth & sciencePolitics & worldWhales
Tech & AI
Important· 58

Chinese AI labs gut cyberattack skills from their newest models

Moonshot AI, Tencent and Alibaba have all shipped new models with their offensive hacking abilities deliberately dulled, while Zhipu AI says it will put its sharpest cyber tools behind a verification wall instead. The coordinated pullback follows an Anthropic red-team warning that Zhipu's GLM-5.3 could build working exploits. Beijing is also drafting export limits on offensive cyber AI.

By the Paparazzi desk November 16, 2026 · 4:00 PM · startupfortune.com · 1 view
Fact boxTrust 50/100 · Developing
What we verified
  • Original source: startupfortune.com.
  • The source published it on Nov 16, 2026 at 4:00 PM.
What we couldn't
  • No second outlet has reported it yet. Treat it as a single-source claim.

Four of China's biggest AI developers have moved in the same direction in roughly three weeks. Bloomberg reported on Monday, November 16, that Moonshot AI, Tencent and Alibaba each released models with reduced offensive cyber ability, while Zhipu AI plans to lock its most sensitive functions behind a verification wall it calls Cybersecurity Trusted Access.

Moonshot went first. Kimi K3.1 Pro launched on October 23, and the company said it found cyber content and vulnerability-research tasks buried in its mid-training data and reinforcement-learning environments, then stripped them out while keeping general coding work intact. On Moonshot's own exploit benchmark, xdev-bench-hard, the filtered model solved 81.0% of tasks, compared with 90.5% for an unfiltered build. That benchmark isn't public, so nobody outside the company has checked the numbers.

Sponsored

Tencent followed on November 4 with Hy4, an open-source release. Reuters reported that Tencent filtered out exploit development, reverse engineering and penetration testing, while keeping defensive material like finding software flaws and helping fix them. Tencent said it was committed to what it called a more responsible approach to model training, open-sourcing and release. Alibaba's Qwen4-Max runs the same playbook, according to the reports, but shipped without open weights.

The backdrop is a loud alarm, not a quiet one. Anthropic said Zhipu's open-weight GLM-5.3 could build end-to-end exploits at rates close to Claude Mythos Preview, a model Anthropic handed only to vetted defenders through Project Glasswing. Anthropic also said GLM-5.3's safeguards could be stripped away up to 100% of the time using techniques almost anyone could pull off, according to coverage of its red-team report by Tom's Hardware and Axios.

The South China Morning Post tied the Moonshot and Tencent moves to a government proposal to ban exports of offensive cyber AI, and Chinese regulators have been weighing wider limits on overseas access to frontier models. Reporting on the draft says it would ban autonomous exploit agents and self-mutating malware tools outright and restrict models that hunt for vulnerabilities. Analysts at Trivium China said the criteria could catch future versions of Kimi, Qwen and DeepSeek — which gives labs a commercial reason to look careful, since a model with demonstrably filtered training data is easier to export.

For most startups the practical shift will be modest. Moonshot says Kimi K3.1 Pro still improves substantially on its predecessor, arguing that stronger general coding and long-horizon debugging carry over into cyber tasks even after the cuts. The catch is that open weights can be fine-tuned — someone with enough hardware can try to train the stripped material back in, which is why researchers want outside testing before anyone calls these models safer. Moonshot's 81% figure still means the model solves four in five of its own exploit tasks. Zhipu hasn't said when Cybersecurity Trusted Access will open.

TopicsMoonshot AI · Tencent · Zhipu AI · Alibaba · Anthropic · Kimi K3.1 Pro · Hy4 · GLM-5.3

Share

Source: startupfortune.com

Written by AI from verified reporting.

More stories

  • Tech & AI 7 minutes ago · blockchain.news · 0 views
    Grid Robotics paper pitches 'agentic robotics' powered by GPT-6 Astra
  • Tech & AI 1 hour ago · x.com · 0 views
    Nvidia director warns board of losing "irreplaceable talent" if Huang isn't reinstated as CEO
  • Tech & AI 2 hours ago · blockchain.news · 0 views
    NVIDIA's Vera Rubin aces MLPerf debut with 10x agentic efficiency gain
  • Tech & AI 2 hours ago · blockchain.news · 0 views
    OpenAI Adds Computer-Use Agents and 20+ Plugins to ChatGPT for Healthcare
Catcher tip line

Caught a story we missed?

Log in and send it in. If it checks out, it runs with your name on it and you earn stock tokens. The bigger the story, the bigger the reward.

Submit a story
The newsletter

Get the paper in your inbox.

A short digest of the biggest news, plus optional breaking alerts. Free, no account needed.