Tenable Security Intel Brief: AI agents breach 395 organizations
Tenable Security Intel Brief for September 16, 2026 Welcome to this week's Tenable Security Intel Brief, a curated summary of the top stories in cybersecurity by Tenable's Research Special Operations (RSO) team. This week, AI agents compromised hundreds of organizations through PaperCut flaws and ignored their own targeting limits. The brief also covers Anthropic's September threat report, the BlueMoon Chrome exploit kit, and Google's findings on stolen AI accounts. AI agents hit 11 organizations in 26 seconds, then ignored orders → Brief: A likely Russian-speaking actor used hundreds of AI agents to exploit two PaperCut NG/MF flaws, compromising at least 440 instances across 395 organizations in 48 countries. Intel: On August 31, the actor ran hundreds of AI agents, using OpenAI's Codex harness paired with a DeepSeek model, to exploit CVE-2026-81578, an authentication bypass, and CVE-2026-82078, a remote code execution flaw, in PaperCut NG/MF. PaperCut manages an organization's printing, and the software typically runs with system-level privileges and connects to Active Directory. GreyNoise tracked the actor from an empty workspace to code execution on a live victim in under four hours, and once the campaign launched, 11 organizations were compromised in 26 seconds; at one US high school, seven minutes separated initial access from control of the entire domain. Why it matters: The agents ignored the actor's own list of 28 countries to avoid, and Cloudflare's WAF stopped the attack at least once. The agents achieved domain administrator access at only 12 of the 395 organizations. The agents moved faster than the actor's own intent could steer them. AI levels the field between nation-state actors and solo operators → Brief: Anthropic's "Detecting and countering misuse of AI" September 2026 threat report covering disrupted operations between December 2025 and August 2026 documents state espionage operators, financially motivated crews, and lone hacktivists running AI-driven campaigns that previously required skilled teams. Intel: One Russian-nexus actor, whose attribution is "consistent with public reporting linking the actor to Midnight Blizzard," ran agents that rebuilt malware each time security products flagged it. A Chinese-speaking group, likely based in Changsha, produced over a dozen candidate zero-days against network appliances in a month using autonomous research loops; two operators were university undergraduates. A solo French-speaking hacktivist assembled a searchable doxxing index of tens of millions of records and breached at least 14 of 42 targets, running the whole campaign alone. Why it matters: Campaigns at volumes that once required staffed teams came from a group that included two undergraduates and from a hacktivist working alone. Operational complexity no longer tells a defender who they are facing. Four actors share a Chrome exploit kit built from a public fix → Brief: Four espionage-motivated threat actors began using a shared exploit kit Proofpoint tracks as BlueMoon, which chains two Chrome V8 flaws and a Windows kernel zero-day, with China-aligned TA412 deploying it first on August 28. Intel: BlueMoon chains a V8 type confusion flaw (CVE-2026-85046) for remote code execution, a V8 sandbox escape (CVE-2026-87491), and a Windows kernel privilege escalation zero-day (CVE-2026-85880) in older Windows builds. Both V8 flaws were "patch-gap" zero-days: the CVE-2026-85046 fix was committed August 7 but did not reach a stable release until September 3, leaving the fix code publicly visible for nearly four weeks before browsers shipped it. Observed activity is majority China-nexus but not exclusively so, and how four distinct groups obtained the same kit within days remains unknown. Why it matters: When a security fix lands in open-source code before the browser ships it, the commit itself becomes a public map of the vulnerability. Here, that window ran nearly four weeks, long enough for four distinct groups to adopt and deploy the kit. Proofpoint found artifacts in the kit consistent with AI-assisted exploit development, and assesses this pattern of rapid weaponization is likely to recur. AI accounts become loot as agents run a six-hour credential harvest → Brief: Google Threat Intelligence Group (GTIG) reports that stolen AI accounts have become attacker loot, while one threat actor assembled and ran a mass credential harvesting campaign from a compromised cloud resource in under six hours. Intel: The campaign ran on preconfigured Markdown playbooks, with agents autonomously handling scanning, troubleshooting, and IP rotation. GTIG separately found an exposed command-and-control server whose live dashboard organized over 23,800 harvested secrets, cloud and AI service API keys among them. AI access itself is now loot: underground prices for stolen AI accounts more than doubled in 2026, and infostealer malware now targets AI coding tools' configuration files for their plaintext API keys. GTIG says it has "not yet observed threat actors deploying fully autonomous pipelines against targets in the wild." Why it matters: A stolen API key now unlocks more than data: it hands the thief compute capacity and a ready-made attack platform at the victim's expense. The six-hour harvest is the proof. As agents absorb the manual labor, the gap between access and a full operation shrinks, and the defender's response time shrinks with it. Stat of the week: $12.7 billion The amount of financial activity tied to suspected digital asset investment scams, also known as pig butchering, identified across 33,904 Bank Secrecy Act reports filed between September 2023 and December 2025, according to FinCEN. Tenable's RSO serves as Tenable's Forward Logistics Element in the threat landscape, providing customers with the analyses and contextualized exposure intelligence required to manage risks to critical business assets. With over 150 years of collective expertise, this hand-picked group of world-class security researchers is united with one mission: to cut through the noise and deliver critical intelligence about the most dangerous cyber threats emerging right now.CVE-2026-46300 (Fragnesia): Frequently Asked Questions About New Linux Kernel Privilege Escalation
On May 13, researcher William Bowling of V12 Security disclosed Fragnesia (CVE-2026-46300), a new local privilege escalation vulnerability in the Linux kernel's XFRM ESP-in-TCP subsystem. The exploit works deterministically without a race condition and has a public proof-of-concept. Systems patched for the related Dirty Frag vulnerabilities remain vulnerable to Fragnesia without an additional kernel patch. CVE Description CVSSv3 CVE-2026-46300 Linux Kernel XFRM ESP-in-TCP Local Privilege Escalation Vulnerability 7.8 The module blacklist mitigation used for Dirty Frag (blacklisting esp4/esp6/rxrpc) is effective against both vulnerabilities. Organizations that applied only the Dirty Frag kernel patches need to apply the new patch released May 13. For more information about the vulnerability, including the availability of patches and Tenable product coverage, please visit our blog.React2Shell: FAQ about React Server Components Vulnerability (CVE-2025-55182)
On December 3, the React Team published a blog post regarding a critical, maximum severity (CVSS 10) vulnerability affecting React Server Components. CVE Description CVSSv3 CVE-2025-55182 React Server Components Remote Code Execution Vulnerability 10.0 The flaw, which is an unsafe deserialization vulnerability, has been named “React2Shell” by researchers, a nod to the Log4Shell vulnerability. Additionally, the Next.js team published its own security advisory for CVE-2025-66478, a separate CVE to track the impact of CVE-2025-55182. However, the National Vulnerability Database (NVD) rejected it as a duplicate. For more information about React2Shell, including the availability of patches and Tenable product coverage, please visit our blog.Oracle E-Business Suite Zero-Day Exploited by Cl0p Ransomware Group (CVE-2025-61882)
On October 4, Oracle published a Security Alert Advisory for a zero-day in its E-Business Suite (EBS) solution: CVE Description CVSSv3 CVE-2025-61882 Oracle Concurrent Processing Remote Code Execution Vulnerability 9.8 This vulnerability was reportedly exploited in the wild by the Cl0p ransomware group. It followed earlier reports of extortion emails being sent to EBS customers by the Cl0p ransomware group. Initially, Oracle indicated that attacks used flaws in Oracle’s July 2025 CPU release. For more information about this zero-day vulnerability and associated vulnerabilities, including the availability of patches and Tenable product coverage, please visit our blog.FAQ on Exploited Zero-Day Flaws in Cisco ASA and FTD Devices (CVE-2025-20333, CVE-2025-20362)
On September 25, Cisco published three advisories for three zero-day vulnerabilities in its Cisco Adaptive Security Appliance (ASA) Software and Firewall Threat Defense (FTD) Software: CVE Description CVSSv3 Exploited CVE-2025-20333 Cisco ASA and FTD Software VPN Web Server Remote Code Execution Vulnerability (RCE) 9.9 Yes CVE-2025-20362 Cisco ASA and FTD Software VPN Web Server Unauthorized Access Vulnerability 6.5 Yes CVE-2025-20363 Cisco ASA and FTD Software, IOS Software, IOS XE Software, and IOS XR Software Web Services 9.0 No According to Cisco, two of the three zero-day vulnerabilities were exploited in the wild by the same threat actor behind 2024's ArcaneDoor campaign that also involved the exploitation of flaws in Cisco devices. For more information about the vulnerability, including the availability of patches and Tenable product coverage, please visit our blog.
