Claude found 493 security flaws in 25 open-source projects

Anthropic says its AI model, Claude, flagged 493 security issues across 25 open-source projects. One of them had gone unnoticed for 21 years. Chris Rohlf of Anthropic published the results, which Techmeme highlighted on October 21, 2026.

What was reviewed

Open-source projects are programs whose code is public, so anyone can read, use and build on them. Many businesses depend on them without knowing it. The review covered more than 5 million lines of code. The languages were C, C++, Java, PHP and Python.

Rohlf framed the work as part of a larger body of evidence. Over the past year, Anthropic has published data showing that leading AI models are capable of finding and exploiting software weaknesses.

Why it matters

A flaw that survives 21 years shows how easily old problems can hide in code that many people trust. If AI tools can find such flaws faster, they may help the people who fix software. The same skill could also help attackers, which is why the topic draws attention.

What is not known yet

The material available does not say how many of the 493 issues were confirmed as real, how many have been fixed, or which projects were involved. It also does not say how serious the flaws were or how the review was run. Those details matter, because a flagged issue is not always a true danger.

What to do

• Keep your software and apps updated, since fixes only help once installed.
• Ask your software vendors how they check their code for flaws.
• Ask whether your vendors use open-source parts and how they track updates to them.