An AI-powered tool has identified 84 previously unknown security flaws in the software that underpins 4G and 5G cellular networks. Researchers at Nanyang Technological University developed the tool, named iFinder, which found these vulnerabilities by analyzing core network software. Of the 84 reported flaws, 83 have been confirmed by developers, and 81 have been assigned Common Vulnerabilities and Exposures (CVE) identifiers.
A particularly severe vulnerability allows an attacker to hijack a subscriber's data session, redirecting their internet traffic to the attacker instead of its intended destination. This exploit leverages a design flaw in how internal network components communicate. In older network architectures, where core components were physically secured, internal links were assumed to be trustworthy. Consequently, the receiving components often accept instructions without rigorous verification. An attacker who gains access to one of these internal links can send a malicious instruction that reuses an existing forwarding rule ID assigned to a legitimate subscriber, but with a higher priority. Since the network prioritizes rules and does not verify the uniqueness of IDs, it adopts the attacker's rule, sending the victim's outbound traffic to the attacker.
This session hijacking attack was successfully demonstrated in a lab environment against OpenAirInterface's (OAI) open-source 5G core. It was subsequently validated on two commercial 5G core networks, including a reproduction within a partner vendor's lab using default settings. One vendor has since patched the issue, which is tracked as CVE-2026-8233. The other major 5G carrier is reportedly still working on a fix. The shift of cellular network cores into cloud environments exacerbates this risk, as misconfigurations can inadvertently expose internal interfaces to the public internet.
The researchers also explored another attack vector: embedding network control messages within the data tunnel of an ordinary phone with a valid SIM card. This technique proved effective against five of the seven open-source cores examined.
iFinder operates through a three-stage AI agent pipeline. The first agent scans code for instances where incoming data is used without proper validation. The second agent cross-references these findings with 3GPP standards documents to determine if the necessary checks occur elsewhere in the communication flow, which often explains seemingly suspicious code. The third agent then attempts to write and execute an exploit against a test network, iteratively refining it based on error logs until it either succeeds or fails. While effective, the tool is not infallible; it caught 15 out of 22 previously known bugs and approximately a quarter of its reports are false positives.
The confirmed flaws have received varying responses from maintainers. Of the 83 confirmed vulnerabilities, 58 have been patched. However, three of the seven projects have not released any fixes, leaving 23 confirmed problems with CVE numbers still unaddressed. This disparity largely reflects differences in maintainer responsiveness and development capacity. For instance, OAI has not responded to the reports, while eUPF acknowledged them but cited resource limitations for implementing fixes. In contrast, Open5GS, SD-Core, and free5GC actively engaged with the researchers and resolved the reported issues.
The researchers ensured that each machine-generated finding was manually reviewed, reproduced, and accompanied by a working proof-of-concept before being reported. This meticulous approach helped maintainers treat the batch of reports as verified security findings rather than raw machine output.
For commercial cores, direct analysis with iFinder was not possible due to the lack of source code access. Instead, the two commercial networks were tested by replaying exploits developed against the open-source versions. Three such exploits were successful: the session hijack on both commercial cores, and two denial-of-service vulnerabilities on one of them, one of which also has its own CVE. The researchers suggest that many commercial 5G core implementations may incorporate or customize open-source stacks, potentially inheriting design assumptions and vulnerabilities from upstream code. While the three confirmed cases in commercial products are a small sample, they serve as supporting evidence rather than a definitive conclusion, indicating a need for further direct analysis of additional commercial cores to determine the full extent of this inheritance.
