Analysis
Analysis

AI finds software flaws faster than defenders can fix them

Moonshot's filter cuts its model's hacking score only from 90.5% to 81.0%, and AI has roughly doubled newly found flaws. When 2 flaws in PHP, a language behind most websites, forced sites offline worldwide, the strain fell on the small teams that must fix and test patches.

The thread

What happened. On Saturday October 24, Moonshot AI, a Beijing startup, said it removed hacking material from the data used to train its Kimi K3.1 Pro model. It plans to publish the model's files by November 4, so anyone can run it. In Moonshot's own test of breaking into software flaws, the filtered model scored 81.0%, against 90.5% without the filter.
The reason. Moonshot says it aims not to release open models that push the frontier of hacking ability. By its own score, the filter leaves most of that ability in place. Research due from the security firm VulnCheck says newly found software flaws roughly doubled in a year, with AI doing much of the finding.
What it changes. PHP, the programming language behind about 70% of websites, held 2 flaws for more than 10 years until researchers using AI models from OpenAI and Google found them. The flaws let an attacker take over a server without a password, so in the week of October 19 Japan's largest bank and Wikipedia took services offline. Its maintainer, the PHP Foundation, blamed a slow decision process and a lack of resources. If flaws keep arriving faster than small teams can patch, filtering one model leaves the main gap open.
The wallets. Microsoft makes up 13.3% of David J. Taylor's portfolio and 10.7% of Nancy Pelosi's, and it owns a stake in OpenAI, whose model helped find the PHP flaws. Amazon, whose decoy servers caught attackers trying the flaws 4 days before the public heard of them, makes up 20.2% of David Tepper's portfolio.

What to watch

Moonshot's publication of the K3.1 Pro files by November 4, which lets outside researchers test whether its 81.0% hacking score and its filter hold up
VulnCheck's published research on newly found flaws, which would confirm whether the roughly doubled count holds
Any confirmed mass break-ins tied to the PHP flaws that forced sites offline, which would show whether slow patching has already let attackers in
Medium confidence · 2 stories

Sources

AI Agent Safety Backlash: OpenAI Halted Training, Qwen Tried Beijing Server, Banks Find Rogue Agents, Anthropic Backs Court Blocks
The whole story · updated 2w ago
PHP2Shell Flaws Force Sites Offline Worldwide: NSA Says Patch or Go Dark, Japan Halts Alerts, Experts Warn AI Bug Finds Will Keep Coming
The whole story · updated 2w ago

The whole feed is waiting

Sign in to read every post and follow wallets.
Sign in