By Hamza Faisal

A security team armed with off-the-shelf AI models has done in 90 hours what used to take researchers nine months — and found 66 previously unknown flaws in the networking code that keeps factories, hospitals and power grids online.

Forescout's research arm, Vedere Labs, built a specialised AI-assisted vulnerability research harness and set it loose on 10 widely used TCP/IP stacks, the foundational software that lets connected devices talk to each other over a network. The results, in a report titled "Finding Critical Infrastructure Bugs with Off-the-Shelf AI," show how dramatically AI is changing the economics of finding security holes — for defenders, and for attackers too.

Nine months of work, compressed into 90 hours

The harness combined publicly available Claude models with existing security tools, and the team applied it to seven stacks Vedere Labs had already audited by hand, plus three it had never examined. The AI produced 362 initial findings; after duplicates were removed and each candidate was tested, 99 vulnerabilities were confirmed.

The whole exercise took about 90 hours — about 10 hours of autonomous AI analysis and around 80 hours of human validation. Manual analysis of the same kind took about nine months during Project Memoria, Forescout's earlier investigation of embedded networking code in 2020 and 2021. It also rediscovered 33 known flaws — about 40% of those in the previously audited stacks.

The cost was equally striking. Automated static analysis across all 10 stacks cost just $935, or about $2.58 per finding, with total model spending of $1,200 to $2,300.

What the AI found — and what it missed

Of the 99 confirmed vulnerabilities, 66 were previously unknown zero-day flaws. More than half of the new discoveries — 52% — were rated high or critical in severity. The team reported 54 of them to the vendors and maintainers responsible, and responsible disclosure continues for the rest. Cybersecurity Dive counted fixes already in place for 24 of the reported flaws.

The stacks examined included Contiki, used in low-power internet-of-things devices; Ethernut, used in Ethernet-based smart devices; lwIP, a lightweight TCP/IP implementation; and wolfIP, used in real-time operating systems in safety-critical devices. The affected software sits inside industrial equipment, medical devices, network equipment and other connected products, and Forescout said past disclosures suggest the flaws likely affect hundreds of products from dozens of vendors.

Some findings have no fix coming at all. Seven vulnerabilities were found in the firmware of a discontinued wireless access point, including a buffer overflow reachable through a firmware upload and hard-coded administrator credentials that cannot be changed — with no vendor left to patch them.

The system was not perfect. Researchers said the harness was strong on memory-safety bugs but misses logic and design flaws, which still need human expertise to catch.

The uncomfortable part: attackers can do this too

Forescout's central point is not just that AI makes defenders faster. The same off-the-shelf models, the same tools and the same tiny budget are available to anyone — including the attackers probing the same code for the same flaws. The barrier to finding vulnerabilities in critical infrastructure has fallen from nine months of expert labour to a few days and a few hundred dollars.

That leaves the organisations running this software with a narrower window than ever. The flaws are in code that ships inside equipment expected to run for years, sometimes decades, and some of it can no longer be patched at all. Finding the bugs faster only helps if the fixing keeps pace — and for the oldest, most embedded systems, that race is already lost.