Skip to content
Brief Security & Trust ·

Forescout used AI to find 66 new flaws in device network code

The security firm says a purpose-built toolkit with Claude models found 99 flaws in about 90 hours.

In brief
  • Forescout found 99 flaws in 10 TCP/IP stacks using Claude models inside a purpose-built toolkit, with researchers validating results, in about 90 hours.
  • The figures leave out the work of building the toolkit. Only five of the 54 flaws reported to makers are fixed so far.

Forescout said it found 99 flaws in 10 TCP/IP stacks, the software that lets devices talk over a network. It used Claude models inside a toolkit built for embedded software, and researchers validated the results. Sixty-six flaws were new. Forescout puts the work at about 90 hours, against about nine months of manual code analysis in its earlier project. That count covers about 10 hours of automated static analysis plus human validation, and it leaves out the work of building the toolkit.

The toolkit first runs code scanners that flag suspect patterns. AI agents then check whether an attacker could reach that code. Last, the software is run, and an agent writes a test that triggers the flaw. Nearly half the 99 flaws were memory corruption in code that reads network traffic. The toolkit missed logic and design flaws, such as guessable random numbers. Forescout reported 54 new flaws to makers. Only five are fixed so far, and the CVE IDs it asked MITRE for are still pending.

Finding flaws in shared device networking code has become cheaper, but fixing them has not. A fix must pass from stack makers to device makers to owners. Budget owners can ask: which devices cannot be patched? Which suppliers commit to support lifetimes? Which software parts lists name these stacks?

A WebPulse Brief: a short report of an important event, written by the WebPulse Newsroom with AI assistance and checked against the reporting below. When there is more to explain, we follow up with a full story. How we use AI.

Reporting: Forescout.