InformationalEcosystem changeMonitoring

CSO Online reports AI-written patches can serve as attack maps

CSO Online reports researchers say AI-assisted patches are often incomplete and help attackers find variants; test AI-written fixes for bypasses.

Published
Exploitation
Not confirmed for any specific flaw; CSO Online reports the Microsoft Edge team cited examples of attackers using Chromium patch gaps

What happened

CSO Online reports, in a feature by Lucian Constantin dated 16 November 2026, that AI is speeding up flaw discovery but that AI-assisted fixes are often incomplete. According to the article, this makes variant analysis more effective for attackers. This advisory summarises that single article. It is a claim-based summary, not an independent verification, and the source text available to us was cut off before its end. CSO Online

What CSO Online says various parties found:

  • vm2: The article says maintainers of the vm2 JavaScript sandboxing library shipped fixes for 34 sandbox-escape vulnerabilities this summer, and that Anthropic’s Claude co-wrote 19 of them. Shayaun Nejad, a lead product security engineer at Rubrik, says he ran variant analysis on a vm2 patch. He says the same bugs kept being rediscovered despite fixes from AI-assisted pipelines. CSO Online
  • Firmware example: OpenAI security researcher Tim de Waal reportedly showed an AI agent patching two flaws in network access point firmware. One patch dropped stack canaries, and the other filtered double quotes but missed single quotes. CSO Online
  • Controlled studies: Per the article, researchers at 1Password’s Off-by-1 Labs found that about half of 6,080 AI-generated patch candidates left at least one attack path open. In the P2Patch benchmark, an attacking agent found a different way to trigger the flaw in 41 of 101 cases. CSO Online
  • Human patches too: CSO Online cites Trail of Bits as saying one in eight fixes for 2,265 reviewed vulnerabilities failed to fully resolve the issue. CSO Online
  • Closed source: The article says Anthropic researchers reported an AI-assisted patch-diffing test that produced a working exploit within an hour of a Firefox fix. It also cites a Fudan University paper that flagged 63 of 76 analysed closed-source patches as incomplete. CSO Online
  • Patch gap: CSO Online reports that the Microsoft Edge Security team warned attackers are increasingly exploiting the gap between a fix appearing in an upstream project such as Chromium and the fix reaching downstream browsers. CSO Online

Not confirmed: the source text provided names no new CVE, no affected product versions, and no specific in-the-wild exploitation of AI-written patches. Whether the findings generalise beyond the cited studies is the claim of the researchers quoted, not something we have verified.

Am I affected?

The article is about a general trend, not one product with a defined version range. It is relevant if your team uses AI agents to write or review security fixes, or if you ship software built on upstream projects whose fixes become public before yours.

  • vm2 affected versions and fixed versions: not confirmed in the source.
  • Whether any specific AI-written patch in your codebase is incomplete: not confirmed; the source offers no way to tell without testing.

What to do now

The source describes problems and a few practices; it does not publish a formal remediation checklist. The steps below follow from what the article reports.

  1. Treat an AI-written security fix as unverified until tested. The article reports that such patches often leave attack paths open, including bypasses of a similar input and the same bug elsewhere in the code. CSO Online
  2. Run variant analysis on your own fixes. Nejad describes pointing an AI agent at a codebase with its historical vulnerabilities as context to hunt for variants. CSO Online
  3. Check that a generated patch does not remove existing protections, as in the stack canary example reported from the [un]prompted conference. CSO Online
  4. If you build on upstream projects such as Chromium, track how long upstream fixes stay public before they reach your product, since the Edge team describes this gap as the window being abused. CSO Online
  5. Keep human review capacity in view. CSO Online reports Meta’s Molnar saying AI-proposed patches still needed human review and that reviewers are overloaded, and that Meta is looking to automate more of the review pipeline. CSO Online

Sources