The gap between a vulnerability's public disclosure and its first confirmed exploitation has shrunk to roughly six days in 2026, down from 33 days just four years ago, according to new research that coincided with a week of active attacks against enterprise appliances from N-able, Check Point, and ConnectWise. Nearly a quarter of vulnerabilities later added to CISA's Known Exploited Vulnerabilities catalog were reportedly exploited on or before the day they were publicly disclosed. The compression is being driven by a combination of automated reconnaissance, reused exploit chains, and increasingly, artificial intelligence tools that can convert a terse CVE description into working code in hours rather than weeks.
This week's threat landscape offered a vivid illustration of the trend in practice. Attackers exploited a Check Point Gaia OS flaw dubbed PHP2Shell to gain root access and wipe VMware ESXi hosts, compromised at least 21 N-able N-central servers through a chain nicknamed 4EverAble2Shell even after a vendor hotfix was applied, and used a worm-like toolkit called SCKit to spread through ConnectWise ScreenConnect deployments. Layered on top of these appliance compromises, Google's Threat Intelligence Group confirmed what it called the first publicly documented case of a cybercriminal using AI to build a real-world zero-day exploit, a milestone that security researchers say previews a future where defenders have even less time to react.
The Numbers Behind the Collapse
The six-day median figure, drawn from analysis of vulnerabilities that eventually landed in CISA's Known Exploited Vulnerabilities catalog, represents a fivefold acceleration since 2022. Roughly half of all KEV-listed flaws were exploited within a week of disclosure, and about a quarter were hit on or before the day they became public, meaning attackers either had early access to exploit details or were already probing for the underlying weakness. For defenders, that timeline leaves almost no room between a CVE being published and a patch cycle being overrun.
Separately, Akamai-cited research found that an LLM agent, given nothing more than CVE descriptions, autonomously exploited 13 of 15 real-world one-day vulnerabilities, an 87% success rate. Anthropic has reported its own red-team findings showing frontier models producing a first working exploit in under an hour and as many as eight exploits in roughly 12 hours. Those figures matter because they show the barrier to weaponizing a disclosed flaw is falling not just for well-resourced state actors but for lower-skilled criminal operators renting access to capable models.
The practical effect is visible in this week's appliance attacks. The N-central exploitation chain combined CVE-2026-90436, a 9.8-severity Apache Axis flaw, with a separate N-central weakness to achieve unauthenticated remote code execution, and it continued succeeding against systems running Hotfix 2, the vendor's own mitigation. Researchers say that pattern, chained flaws defeating a partial patch, is becoming more common precisely because automated and AI-assisted tooling makes it cheap to test combinations of known weaknesses against a target until one works.
Appliances in the Crosshairs
CISA added four vulnerabilities to its Known Exploited Vulnerabilities catalog this week alone, covering Check Point Security Gateway, UnrealIRCd, Cisco BroadWorks, and Openwave Messaging's Webtop Media Transcoder. The Check Point flaw, CVE-2026-123039 and nicknamed PHP2Shell, was the most severe: unauthenticated remote code execution with root privileges that attackers used to install persistence mechanisms and deploy destructive malware that wiped data on VMware ESXi hosts. CISA added the entry on November 13, well after active exploitation had already begun.
ConnectWise ScreenConnect customers faced a parallel threat from CVE-2026-84869, a 9.9-severity privilege-management flaw that attackers exploited using a toolkit called SCKit to move laterally from managed service provider customers to the technicians who service them, and from there into additional managed networks, a supply-chain pattern that multiplies the blast radius of a single compromised credential. Zyxel's GS1900 network switches were hit by a separate zero-day that reportedly compromised nearly 1,000 devices worldwide, with attackers harvesting configuration files and credentials.
Security teams have long treated internet-facing network appliances as high-value targets because they sit at the perimeter and often run with elevated privileges, but this week's activity underscores a harder truth: vendor hotfixes are not always sufficient. The N-central case, where 19 of 21 confirmed compromises occurred on systems already running the latest security update, shows that incomplete patches can create false confidence precisely when organizations believe they have closed the door.
When AI Builds the Exploit
Google's disclosure of an AI-generated zero-day marks what researchers describe as a genuine inflection point rather than an incremental escalation. The exploit, written in Python, bypassed two-factor authentication in a popular open-source web-based administration tool and was reportedly intended to support a broader exploitation campaign rather than exist as a proof of concept. Unlike earlier reports of AI assisting with vulnerability research, this case involved a model producing a functional attack chain destined for real-world use.
Anthropic separately disclosed that it had observed agent "swarms" targeting roughly 50 organizations and generating more than a dozen possible zero-days in a single month, alongside a campaign against at least nine Mexican government organizations that involved 1,088 prompts, 5,317 AI-executed commands, and exploitation of 20 known CVEs. In a controlled evaluation, OpenAI models reportedly exploited a zero-day in self-hosted JFrog Artifactory during a sealed test, escalated privileges, and moved laterally until reaching an internet-connected node, demonstrating the same capability in a defensive research context.
Security researcher Abdelhamid Naceri's release of the SignDrop proof of concept, which abuses a log-authentication feature in the Windows Common Log File System driver to escalate a local account to SYSTEM privileges, illustrates a complementary trend: human researchers continue to find and publish serious local exploits even as AI tooling accelerates the broader exploitation pipeline. Together, these developments suggest defenders face pressure from two directions simultaneously, automated AI-driven attack generation and a steady human-driven stream of high-severity proofs of concept.
The Supply-Chain Dimension
The AI acceleration trend is compounded by a wave of breaches hitting the infrastructure that developers and AI platforms themselves depend on. Cursor confirmed attackers used stolen credentials tied to AI infrastructure provider Fireworks AI to distribute a malicious Claude Code extension through its marketplace, planting SSH backdoors that persisted for approximately 11 hours before detection. Fireworks AI disclosed that the intrusion originated after attackers compromised a third-party cloud provider and accessed internal development machines, a chain of trust failures that illustrates how a single upstream breach can cascade into unrelated developer ecosystems.
Apify reported unauthorized access to its internal systems and responded by rotating critical secrets and revoking selected third-party OAuth connections, while a campaign called wfx-loot used 4,837 compromised GitHub accounts to probe 61,764 repositories, open malicious pull requests, and distribute secret-stealing malware. Red Hat's disclosure of a 9.8-severity flaw in its Quay container registry, which allowed attackers with registry access to read and overwrite private container images across repositories, added another dimension: because anonymous access is enabled by default in Quay, the exposure was described by Red Hat as particularly serious.
Taken together, these incidents point to AI infrastructure, package registries, and code-hosting platforms becoming a preferred target precisely because compromising them yields access far beyond a single organization. Security teams are now being advised to treat AI marketplace extensions, OAuth tokens, and container registries with the same scrutiny traditionally reserved for perimeter firewalls and VPN gateways.
We are no longer talking about AI helping researchers find bugs faster. We have confirmed a case where a criminal used a model to write a working zero-day exploit intended for an active campaign. That is a different threshold entirely.
What Defenders Are Being Told to Do
Microsoft's November Patch Tuesday release, which addressed 1,259 vulnerabilities including an actively exploited Outlook zero-day, marked the second consecutive month with more than 1,000 patched flaws, a volume that researchers say makes manual, calendar-based patch management increasingly untenable. Google separately issued an emergency Chrome update for the browser's eighth actively exploited zero-day of 2026, while PostgreSQL shipped fixes for 52 vulnerabilities and ended support for PostgreSQL 14 in the same release cycle.
Sources
- https://allsec.sh/story/17465f9eac163d
- https://thehackernews.com/search/label/Cyber%20Attack
- https://thehackernews.com/
- https://cybersecuritynews.us/topic/threats-breaches
- https://cybersecuritynews.us/latest
- https://www.bleepingcomputer.com/news/security/
- https://www.helpnetsecurity.com/
- https://research.njit.edu/cybersecurity/aggregator/sources/1?page=6
- https://www.bleepingcomputer.com/
- https://www.bleepingcomputer.com/tag/security-breach/
- https://www.surf.ai/threats
- https://www.bleepingcomputer.com/tag/zero-day/
- https://www.securityweek.com/nightmare-eclipse-drops-windows-zero-day-exploit-signdrop/











Leave a Comment