Skip to content
NewA calmer, faster ZeroHour

AI-written patches leave variant attack paths, researchers say

Original title: AI-written patches are becoming attack maps for hackers

CSO Online ()9 min read1
AI summary · x-ai/grok-4.7Summaries by Vani AI

In cited tests, about half of 6,080 AI patch candidates left an attack path, and a follow-on agent bypassed fixes in 41 of 101 cases.

At the [un]prompted conference, Rubrik's Shayaun Nejad said an AI agent found a bypass for an AI-assisted vm2 sandbox-escape fix after maintainers shipped patches for 34 issues, 19 co-written by Claude. Other talks and studies described incomplete AI patches, including OpenAI work on access-point firmware and a 1Password Off-by-1 Labs run on six known bugs. Anthropic's Glasswing ledger and comments from Bank of America and Meta were cited as evidence that AI finding outpaces human review and patching.

Unconfirmed: Anthropic's 94.2% true-positive rate and several conference figures are claims by the speakers or firms cited, not independently confirmed here.

Key facts
  • vm2, with 1 million downloads a week, shipped fixes for 34 sandbox escapes; Claude co-wrote 19.
  • Off-by-1 Labs reported about half of 6,080 patch candidates left at least one attack path.
  • In the P2Patch benchmark, a second attack succeeded against 41 of 101 patched vulnerabilities.
  • Trail of Bits found one in eight of 2,265 human-submitted fixes failed to fully resolve the issue.
  • VulnCheck's reading of Glasswing counted 217 fixes among 7,349 public flaws; Anthropic claims a 94.2% true-positive rate.
Security assessment · AI, from the article

Impact: Incomplete patches can leave the original or variant attack path open

ProductAffected
vm2—
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

From the original article

9 min read at csoonline.com

This summer, the maintainers of vm2, an open-source JavaScript sandboxing library with 1 million downloads per week, shipped fixes for 34 separate sandbox-escape vulnerabilities and Anthropic’s Claude co-wrote 19 of them. Weeks later, a security researcher pointed an AI agent at one of the advisories and asked for a way around the patch. The agent delivered.

A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.