CryptoSlate
Forescout researchers used commercially available AI to find 66 previously unknown vulnerabilities in embedded networking software, part of a batch of 99 validated flaws across 10 TCP/IP stacks that let connected devices communicate over networks. The system paired conventional code analysis with Anthropic's Claude models, examining both open-source code and proprietary binaries before dynamic testing confirmed each find.
The economics are the headline for security teams: roughly 10 hours of automated analysis and 80 hours of dynamic validation, with static analysis costing about $935 in model spending and total model spend estimated at $1,200 to $2,300. Forescout's earlier Project Memoria work took about nine months of manual code analysis, though the comparison spans different target sets and excludes the cost of building the system itself.
Why it matters
The approach is not a replacement for expert research. It rediscovered only 40% of known issues in a previously studied target set and was weaker at finding implementation-specific logic and design flaws. But the cost curve matters: vulnerability discovery at this price point changes the balance between defenders and attackers, since the same tooling is available to anyone.
The findings also expose a patching gap. Siemens' Oct. 13 advisory covers flaws in APOGEE, TALON and Desigo building-automation products where an unauthenticated remote attacker could read device memory or trigger denial of service. Updates exist for APOGEE PXC and TALON variants, but several listed legacy APOGEE MBC/MEC and Desigo products have no planned fix and require mitigations such as disabling FTP/web services or network segmentation.
Market impact
An open-source fix for the lwIP stack landed Oct. 26, addressing an MQTT parsing flaw that could let a malicious broker freeze a device's network services. Disclosure remains unfinished, with three proprietary stacks still under investigation and one vendor, Ethernut, yet to respond. Operators should cut unnecessary service exposure, isolate unpatchable devices and apply available fixes. Faster discovery only helps if vendors actually ship patches.
Frequently asked questions
-
How many vulnerabilities did Forescout's AI-assisted research find?
The team validated 99 vulnerabilities across 10 TCP/IP stacks, including 66 previously unknown flaws. It reported 54 of the new findings to vendors or maintainers because they still affected current versions.
-
How much did the AI-assisted vulnerability research cost?
Static analysis cost about $935 in model spending, with Forescout estimating total model spend at roughly $1,200 to $2,300. That covers applying an already-built system, not the expert research and setup behind it.
-
How does the AI approach compare to manual security research?
Forescout logged about 10 hours of automated analysis and 80 hours of dynamic validation, versus roughly nine months of manual code analysis in its earlier Project Memoria work. The comparison spans different target sets, so it is not a controlled benchmark.
-
Which devices have no fix planned for these flaws?
Siemens lists updates for affected APOGEE PXC and TALON variants, but several listed legacy APOGEE MBC/MEC and Desigo products have no planned fix. Siemens recommends issue-specific mitigations, including disabling FTP/web services or placing controllers behind network segmentation.
-
What are the limits of using AI to find security vulnerabilities?
The system rediscovered only 33 of 82 known issues, about 40%, in a previously studied target set, and was less effective at finding implementation-specific logic and design flaws. It still required 80 hours of human validation.
WatcherGuru
CoinTelegraph
Lookonchain
Crypto News
WuBlockchain
Altcoin Daily
CoinDesk