Our weekly round-up of significant threats and actively exploited vulnerabilities. It is awareness content, not a monitored threat intelligence service: always check vendor advisories against your own environment.
This week at a glance
Attackers moved quickly against two newly disclosed flaws: a path traversal bug in Atlassian’s self-hosted Data Center and Server products, exploited within days of a public proof of concept, and a chain of flaws in AhsayCBS backup servers, which are widely used by managed service providers. CISA also added five older flaws to its Known Exploited Vulnerabilities (KEV) catalogue after the UK’s NCSC and international partners called out China-based Integrity Tech. On the threat side, the FBI warned that FortiBleed credential attacks on FortiGate firewalls are still under way, and a large GitHub Actions campaign stole cloud and AI keys from thousands of repositories.
Actively exploited vulnerabilities
Atlassian Data Center and Server (CVE-2026-21589)
A critical path traversal flaw (CVSS 4.0 score 9.3) lets unauthenticated attackers read files from self-hosted Jira, Jira Service Management, Confluence, Bitbucket, Bamboo, Crowd, Crucible, and Fisheye. Reading Crowd’s configuration file can let an attacker create a Jira administrator account. Atlassian disclosed the flaw on 5 October and exploitation was reported from 7 October, shortly after a public proof of concept appeared. Atlassian Cloud is already patched. Fix: upgrade to the fixed release for each product (for example Jira 9.12.40, 10.3.26, or 11.3.12, and Confluence 9.2.26 or 10.2.19); until you can, apply Atlassian’s WAF or RewriteValve workaround and restrict internet access, then rotate Crowd application credentials.
Source: Atlassian security advisory
Ahsay AhsayCBS (CVE-2026-105134, CVE-2026-105133)
Huntress reported on 8 October that attackers are chaining an authentication weakness (CVE-2026-105133) with an unauthenticated command injection flaw in the replication receiver (CVE-2026-105134, CVSS 4.0 score 9.3) to run code as SYSTEM on AhsayCBS backup servers, then installing web shells and a cryptocurrency miner. Exploitation began on 7 October, three days after the flaws were published, and hit at least five organisations. AhsayCBS is mainly used by managed service providers and backup resellers. Huntress says version 10.3.4 is still vulnerable, and no fixed version has been confirmed. Fix: take the AhsayCBS management console off the internet (trusted IP addresses or VPN only), hunt using Huntress’s indicators, and re-image any compromised server.
Source: Huntress
The next five are older flaws that CISA added to the KEV catalogue on 8 October because they are used in the activity described in the joint advisory on Integrity Tech (see below). They matter most if you still run legacy or unmanaged internet-facing systems.
ProFTPD (CVE-2015-3306)
The mod_copy module in ProFTPD 1.3.5 lets unauthenticated users copy files on the server using the SITE CPFR and SITE CPTO commands, which can lead to remote code execution (CVSS 10.0). Fix: upgrade to a supported ProFTPD release or disable mod_copy, and retire internet-facing FTP where you can.
Source: NIST National Vulnerability Database
ONLYOFFICE Document Server (CVE-2021-3199)
A path traversal flaw in the upload function of ONLYOFFICE Document Server before 5.6.3 can lead to remote code execution (CVSS 9.8). Fix: upgrade to a current Document Server release and avoid exposing it directly to the internet.
Source: NIST National Vulnerability Database
Apache Struts 2 (CVE-2016-3081)
When Dynamic Method Invocation is enabled, affected Struts 2 releases allow remote command execution (CVSS 8.1). Fix: upgrade to a supported Struts release (the original fixes were 2.3.20.3, 2.3.24.3, and 2.3.28.1) and disable Dynamic Method Invocation.
Source: NIST National Vulnerability Database
Strapi (CVE-2023-22894)
A user with access to the Strapi admin panel can abuse query filters to expose password hashes and password reset tokens, and the flaw can be chained with CVE-2023-22621 for code execution. Fix: upgrade to Strapi 4.8.0 or later and rotate admin passwords and reset tokens.
Source: NIST National Vulnerability Database
ISC BIND 9 (CVE-2015-5477)
A crafted TKEY query can crash the BIND name server, causing a denial of service. Fix: make sure no legacy BIND 9.9 or 9.10 servers remain (the original fixes were 9.9.7-P2 and 9.10.2-P3) and run a supported release.
Source: NIST National Vulnerability Database
Notable threats and campaigns
FortiBleed: FBI warns attackers are locking out FortiGate administrators
On 6 October the FBI and the US Secret Service warned that the FortiBleed operation is still active, affecting FortiGate devices worldwide (194 countries). Attackers log in to FortiGate firewalls and SSL VPNs with credentials leaked earlier this year or taken from infostealer logs, crack password hashes offline, create rogue administrator accounts, and lock out the real administrators before selling the access to ransomware affiliates, including INC/Lynx and Payload. The FBI warns that patching and resetting passwords alone may not be enough: end all administrator and VPN sessions, review administrator accounts and API keys, use phishing-resistant MFA, and take management interfaces off the internet.
Source: FBI and US Secret Service advisory
GhostAction: GitHub Actions workflows steal cloud and AI keys
Socket, StepSecurity, and GitGuardian reported a new wave in which more than 500 compromised GitHub accounts committed a fake “security audit” workflow (security-audit.yml or github_actions_security.yml) to tens of thousands of repositories. The workflow sends GitHub Actions secrets, along with cloud and AI service keys found in the repository history, to an attacker’s server. Maintainers’ tokens were probably taken from infostealer logs. Search your organisations for these file names, remove them from every branch and fork, rotate any exposed secrets, and turn on secret scanning with push protection.
Source: Socket
FakeGit returns with 17,610 malicious GitHub repositories
Apiiro reported that the FakeGit campaign has pushed 17,610 lure repositories, more than 13,000 of them in 34 hours, including about 800 posing as AI skills or MCP servers. A “Download” ZIP linked from each README installs the SmartLoader loader and then the StealC infostealer, and at least 700 of the accounts appear to belong to hijacked legitimate developers. Tell developers to install tools and MCP servers only from official registries, and treat any GitHub account used on a machine where SmartLoader ran as compromised.
Source: BleepingComputer
Also reported this week
- Integrity Tech joint advisory: the NCSC and partners in Australia, Canada, Japan, New Zealand, Spain, and the US called out China-based Integrity Tech, whose tools have been used against targets including airports in Japan and Poland and organisations in the US and Taiwan, and the FBI seized domains behind its scanning and phishing tools (NCSC).
- TensorLake npm compromise: version 0.5.144 of the tensorlake package carried a self-spreading credential stealer that can wipe data if its stolen token is revoked, so remove its persistence before rotating credentials (Socket).
- Fake AI tool sites steal advertising accounts: fake ChatGPT, Gemini, Claude, and Perplexity sites use fake sign-in pop-ups and a live operator to capture passwords and MFA codes from advertising agency staff (BleepingComputer).
- Malvertising for a fake Claude download: Google ads redirected through Bing lead Mac users searching for AI tools to a page whose copy button swaps in a malicious terminal command (BleepingComputer).
- Power BI links deliver rogue ScreenConnect: phishing emails link to attacker-made Power BI reports that download unauthorised ScreenConnect remote access clients (Huntress).
- Midnight Mimosa: ad fraud and proxy malware comes preinstalled in the firmware of low-cost MediaTek Android phones, detected in more than 150 countries, with France, Italy, Germany, and Spain among the most affected alongside Mexico, the US, and Brazil (Bitdefender).
- Qilin arrest: Japan detained and extradited to Germany a suspected core member of the Qilin ransomware operation, which has hit more than 2,350 organisations in 62 countries, including Japan, the US, and Australia; the group remains active (BleepingComputer).
Other patches worth knowing
- Citrix NetScaler ADC and Gateway (CVE-2026-107406): a critical memory overflow (CVSS 4.0 score 9.5) in SAML configurations, fixed in 14.1-73.46 and 13.1-64.29; not reported as exploited yet, but NetScaler flaws are usually attacked quickly (Citrix).
- SonicWall SMA 1000 (CVE-2026-102255 and three others): a maximum-severity pre-authentication server-side request forgery flaw, fixed by hotfix; one researcher reported exploitation attempts on 9 October, but SonicWall had not confirmed exploitation (SonicWall).
- Android October 2026 bulletin: fixes for 25 vulnerabilities, seven of them rated critical (SecurityWeek).
- LMCache (CVE-2026-105192): an unpatched remote code execution flaw with a public proof of concept in this open-source LLM cache, so keep it off untrusted networks (JFrog).
- AnyDesk for Linux: a working root exploit for version 8.0.2 has been published; update to 8.0.3 or later (The Hacker News).
Practical steps to consider
- Check that self-hosted collaboration, backup, file transfer, and document servers are patched and are not exposed to the internet unless they need to be. ISO/IEC 27001 control A.8.8 and Cyber Essentials both expect high-risk patches within 14 days.
- Review administrator accounts, API keys, and active sessions on firewalls and VPNs, and use phishing-resistant MFA for remote and administrator access (control A.8.5).
- Treat code repositories and build pipelines as part of your supply chain: turn on secret scanning, limit who can add workflows, and control which packages and tools developers install (controls A.5.21 and A.8.28).
This round-up is for general awareness only and is not a monitored threat intelligence service. If you would like help reviewing your vulnerability management or incident readiness, contact us.
