Quick Take
- What: Forescout Research published findings from an AI-assisted pipeline that discovered 99 vulnerabilities in 10 embedded TCP/IP stacks, 66 previously unknown
- How it works: The pipeline chains Semgrep and Joern static analysis with Claude Sonnet 4.6 and Opus models for triage and verification, achieving 11% to 70% precision depending on the stack
- Who is affected: Embedded and critical infrastructure devices running Contiki, FNET, PicoTCP, Nucleus NET, wolfIP, lwIP, Ethernut and three closed-source stacks
- Timeline: The 90-hour run included roughly 10 hours of AI analysis and 80 hours of human validation
- What to do: Prioritize embedded device risk by reachability and exploitability, isolate devices that cannot be patched, and restrict exposed services such as FTP, Telnet, IPv6 and web admin interfaces
In plain English
Researchers at Forescout used AI models from Anthropic to help find security flaws in the networking code that runs inside embedded devices like industrial controllers and wireless access points. The AI system found 99 bugs in about 90 hours, work that would normally take nine months. However, the AI only did about 10 hours of the actual analysis. The rest of the time was spent by human experts checking whether each finding was real, because the AI was wrong 30% to 89% of the time depending on which code it looked at. Twenty-four of the bugs have been fixed, but 29 others in proprietary code are still being investigated by vendors.
What you should do: If you use embedded devices in your home or business, you do not need to do anything right now, but ask your IT team or vendor whether your devices are affected and whether patches are available.
What happened
Forescout Research published *Finding Critical Infrastructure Bugs with Off-the-Shelf AI* on November 12, 2026. Stanislav Dashevskyi, Francesco La Spina and Daniel dos Santos applied an AI-assisted pipeline to 10 embedded TCP/IP stacks, seven of them revisited from the team’s earlier Project Memoria work. The run confirmed 99 vulnerabilities, 66 of which were previously unknown.
The flaws sit in embedded networking code: Contiki, FNET, PicoTCP, Nucleus NET, wolfIP, lwIP and Ethernut, plus three closed-source stacks Forescout labels S1 to S3. Twenty-nine findings in those proprietary stacks were still under vendor investigation at publication. Forescout reports that seven flaws sat in a stack used by an unsupported wireless access point, including hardcoded administrator credentials.
Forescout rates 52% of the 66 new findings critical or high severity. The team reported 54 of the 66 new findings: 20 in open source stacks and 34 in proprietary ones. Twenty-four are already fixed, across Contiki, FNET, PicoTCP, Nucleus NET, wolfIP and lwIP. The Ethernut maintainer had not responded to the disclosure.
How it works
The pipeline chains ordinary static analysis with model-driven triage rather than replacing either. Semgrep and Joern, running 32 custom rules, narrow the code to candidate sites, and Ghidra handles the closed-source stacks. Claude Sonnet 4.6 drives the first three stages at high and max reasoning, and Opus models between 4.6 and 5 verify candidates dynamically.
Forescout reports precision between 11% and 70% depending on the stack, so human validation still consumed most of the 90 hours. Roughly 10 of the 90 hours were AI analysis, and the remaining time was human validation. Daniel dos Santos states that AI did not remove the need for expert vulnerability researchers but dramatically compressed the work.
Forescout reports model spend of $935 for static analysis, with an estimated total of $1,200 to $2,300.
What you should do
Prioritize embedded device risk by reachability, exploitability and operational impact, and restrict exposed services such as FTP, Telnet, IPv6 and web admin interfaces. Isolate devices that cannot be patched, and build support lifetimes, disclosure policies and SBOMs into procurement.
Forescout advises maintainers to rerun variant analysis over previously audited code and to expand triage capacity for a higher volume of AI-generated reports.
What it means
The research demonstrates that AI models can accelerate vulnerability discovery in embedded systems, but not replace human expertise. The team estimates the same work would have taken nine months by hand. However, the reported precision of 11% to 70% means that most candidate findings still required a researcher to confirm or reject them.
Forescout requested CVE IDs for the open source findings, but the report states that none had been assigned when it was published. Legacy exposure remains a concern: seven flaws sat in a stack used by an unsupported wireless access point, including hardcoded administrator credentials, which the vendor may never patch.
