Exploited at the edge: what four years of confirmed attacks say about patching priorities

After the backlog years, CISA added 187 entries in 2023 and 186 in 2024. Additions then rose to 245 in 2025, an increase of 32%, and reached 295 in 2026 by 10 November, already more than in any full year since the backlog.

The 913 entries cover 222 vendors. Microsoft is the largest single vendor with 143 entries (15.7%), which reflects how widely its software is deployed as much as its security. The long tail matters more for most organisations: 196 of the 222 vendors have five or fewer entries each, and 119 have only one, so an inventory limited to major vendors will miss much of the list.

A growing catalogue does not by itself prove that exploitation is increasing. Additions also depend on how much evidence reaches CISA and how fast it processes it. What it does show is that the volume of confirmed, actionable exploitation that defenders need to respond to has grown.

Of the 739 vulnerabilities first published in 2023 or later, CISA confirmed exploitation of 193 (26.1%) on or before the day they appeared in the NVD, and of 388 (52.5%) within seven days. The median gap was six days.

The pattern differs sharply by product. Operating systems and mobile platforms have a median of zero days: 52.8% were exploited at or before publication. These are typically zero-days fixed in emergency updates, so the defence is fast automatic updating rather than prioritisation. Edge devices have a median of five days, with 57.7% confirmed within a week. Server and web software is slower, with a median of 16 days, which gives organisations a short but real window to act.

Confirmation by CISA comes after exploitation begins, so these figures understate how fast attackers move. A remediation target measured in weeks is behind the threat for most of this list.

We classed 196 entries (21.5%) as network edge and remote-access products: firewalls, VPN and secure-access gateways, application delivery controllers, routers, SD-WAN, email security gateways and remote-support tools. Their share rose from 16.0% of additions in 2023 to 24.4% in 2026.

These devices are attractive because they sit on the internet, hold credentials and network access, and often cannot run endpoint security software, so compromise can go unnoticed. Exploitation is concentrated: Cisco (37 entries), Fortinet (20), Ivanti (16), Citrix (13), Palo Alto Networks (12) and SonicWall (11) account for 109 of the 196 edge entries (55.6%). This partly reflects their large installed base, and we do not suggest that their products are less secure than others.

CISA records known use in ransomware campaigns for 21.4% of edge entries, compared with 9.9% for operating systems and 4.9% for browsers and desktop applications. Edge devices are a common way into a network for ransomware groups.

The type of weakness exploited depends on the type of product. In operating systems and browsers, memory-safety errors such as use-after-free and out-of-bounds writes account for 46.9% and 45.1% of entries. These are hard to eliminate in code written in C and C++.

Edge devices and server software are different. Injection flaws, which let attacker input run as commands or code, and authentication or access-control failures, which let attackers skip the login, together account for 55.1% of edge entries and 50.0% of server entries. The single most common weakness in the whole window is OS command injection (CWE-78, 64 entries). Server software also shows unsafe deserialisation (10.9%), which is almost absent elsewhere.

These classes are well understood and largely preventable through secure design, such as parameterised commands, strong input handling and authentication enforced on every request. Their persistence in security products themselves is the clearest case for asking vendors about their development practices before buying.

Not everything in the catalogue is new. Of the 913 entries added since 2023, 174 (19.1%) concern vulnerabilities published before 2023, the oldest in 2004. Measured by age at the time CISA added them, 198 (21.7%) were more than a year old and 123 (13.5%) more than three years old. 40 of the year-old entries are edge devices.

Attackers return to old flaws because they still work: devices that were never updated, software that is out of support, and systems missing from inventories. CISA has confirmed known ransomware use for 26 of the 198 entries that were over a year old.

Vulnerability programmes that report only on recent issues will understate this risk. An old vulnerability newly confirmed as exploited should be treated with the same urgency as a new one.

The first AI application platform entered the catalogue in May 2025, when CISA added a Langflow vulnerability. By 10 November 2026 there were 14 entries for AI, machine-learning and data-science tooling, 13 of them added in 2026: Langflow (6, including IBM’s distribution), LiteLLM (3), and one each for MLflow, Ray, GPUStack, Marimo and the n8n workflow platform, which is widely used to build AI agents.

These tools are often deployed quickly by engineering or data teams, exposed to the internet for convenience, and connected to model API keys, data stores and other systems. That makes them valuable to attackers and easy to miss in a central asset inventory.

Fourteen entries is too few to call a trend, which is why we rate this finding low confidence. It is an early signal that these platforms belong in vulnerability management on the same terms as any other internet-facing software.