Forescout’s AI Harness Turned up 99 Iot Stack Bugs in 90 Hours

AI has massively sped up the job of tracking down vulnerabilities in the embedded software that critical infrastructure relies on, security company Forescout announced on Thursday.

Using a specialised AI harness, its researchers turned up 99 vulnerabilities over 90 hours, which by their estimate would have been nine months of work without the technology, Forescout’s report says. This work is the newest sign that AI is changing the way weaknesses in critical tech are uncovered, both for good and for bad.

Ten Stacks, Two Kinds of Bugs

Forescout’s researchers set the harness to work on 10 widely used TCP/IP stacks, seven taken from the firm’s earlier work on TCP/IP vulnerabilities plus three it hadn’t examined.

Most of the stacks are open source, and they included Contiki, found in low-power IoT devices like street lights; Ethernut, which runs in smart devices built on Ethernet; lwIP, another lightweight implementation of TCP/IP; and wolfIP, which turns up in the real-time OS software of devices where safety is critical. The harness, which used Claude models, found 33 vulnerabilities again in the stacks Forescout had already tested, plus 66 new ones in those stacks and in the ones it hadn’t tested before. About half were memory corruption bugs, 27% involved integer problems, 16% were input validation failures and 8% were control flow or logic flaws. According to the researchers, this shows the harness is good at finding memory safety bugs in code that parses protocols but misses logic and design bugs. They said that wasn’t surprising, since they had used static analysis rules focused on generic code patterns to direct the harness.

Memory safety issues lend themselves better to rules like these, the researchers said, while logic bugs are specific to each implementation and much harder to catch with a sensible balance between false positives and false negatives. The harness came across the logic bugs by accident, the report says.

Cheaper Than a Human Bug Hunter

The AI models found bugs not only far faster than a person could, according to Forescout, but also far more cheaply. Forescout reckons it spent $1,200 to $2,300 on tokens, which it put at roughly a hundredth of what an experienced bug hunter would earn going through code by hand.

Of the 66 new vulnerabilities, Forescout passed 54 on to the people maintaining the products involved (it held back the rest, as it thought they were no longer present in the latest versions). So far 24 of those have been patched, and maintainers are checking another 29. The final bug sits in Ethernut, and the developer there didn’t reply to the disclosure. The results show both AI’s ability to spot flaws and how exposed critical technology still is, Forescout said. Since the new bugs are in software that makers of IoT and industrial devices favour, they could hit hundreds of products in a number of industries, the report said.

Old Gear and Missed Flaws

What worries the researchers most is how many vulnerabilities are turning up in discontinued kit that is still in wide use. Inside the networking stack of a wireless access point that is no longer supported, the harness spotted seven vulnerabilities, hardcoded administrator credentials among them. Forescout warned that there are plenty of unsupported legacy products like that access point around, and that AI-assisted research can now cheaply and quickly find dangerous flaws that no vendor will ever patch. Forescout was also uneasy about how many bugs its staff turned up in projects that had already been gone over carefully by them and by other experts. The researchers said they were surprised that so many basic vulnerabilities in core networking protocol code had lasted for years in targets they were re-examining, and admitted that those bugs had been missed by them, by other researchers and by the vendors of the software too.

Subscribe to our weekly news digest.

Join now and become a part of our fast-growing community.

Invalid Email Address
Thanks for Subscribing!
We'll be sending you our best soon!
Something went wrong, please try again later