Abliterated open-weight models sold through hosted APIs lower attackers' skill barrier and expose defenders to unvetted inference and data risks.
LevelBlue found 290 GitHub repositories referencing abliteration, nearly 9,000 Hugging Face checkpoints, and commercial gateways selling reduced-refusal models per token. The risk lies in unapproved inference services and coding agents with credentials and production access, not the models themselves.
Phishing emails impersonating Microsoft outage notifications push Microsoft 365 users to credential-harvesting pages hosted on Oracle Cloud, bypassing Abnormal Security, Mimecast, and Microsoft ATP.
Spoofed "Outage Notification" emails claim held messages will be deleted within 24 hours. The landing page rejects the first credential submission, then redirects to the genuine Microsoft login on the second attempt to conceal the theft.
Canadian Cyber Centre lists November 2026 patches for Dell, IBM, Red Hat, and Qualcomm, including IBM webMethods information-disclosure flaw CVE-2026-123628.
No exploitation is reported. Affected ranges include Dell Ansible Storage for PowerFlex before 3.2.0, IBM ContextForge MCP Gateway up to 1.1.0, Red Hat Enterprise Linux 6/7/10, and Qualcomm products patched in the November 2026 bulletin.
Typosquatters registered one-letter lookalikes of 25 brand-protection domains within 24 hours, serving parked pages and scam ads to users who mistype campaign URLs.
Attackers monitored public zone files and RDAP records for new registrations at CSC and MarkMonitor, then registered variants through Cloud Yuqu LLC and Chengdu West. A DNS change could redirect the lookalikes to credential-harvesting forms.
Revived Retadup USB worm C2 servers deliver CoffeeLoader to Windows systems via infected drives, using a GPU-based Armoury packer and OpenGL decryption.
Retadup's 2019-takedown domains were re-registered in 2024 and now issue Base64/hex download commands. An NSIS installer drops a patched 7z.dll whose shellcode injects into dllhost.exe or svchost.exe and fetches the main body over HTTPS.
Socket reported that attackers poisoned a developer library and desktop application, while Huntress and Microsoft described a worm that used compromised ScreenConnect environments to reach customer endpoints. Check Point also reported destructive intrusions through its Mobile Access gateways, adding a local privilege-escalation flaw to previously reported PHP2Shell exploitation.
Investigators documented espionage against Exchange servers and government web applications, a post-compromise ESXi implant, and attacks on exposed AI computing infrastructure. Reports on offensive AI included a Cairn-assisted intrusion campaign and additional findings about ARTEX; separate Microsoft experiments examined how poisoned training data or modified model weights could subvert coding agents.
Node2Shell (CVE-2026-111497), a Brotli decoder RCE in Node.js, is exploited in the wild while AI tools leave multiple unmanaged Node runtimes unpatchable on endpoints.
AI coding agents, MCP servers, and Electron apps install four to five independent Node runtimes per endpoint, most invisible to inventory. Node 18 and 20 are end of life; npx -y <server>@latest executes unpinned registry packages with user privileges at every agent launch.
China-nexus and criminal actors use ARTEX, an open-source AI penetration-testing framework released July 2026, to automate reconnaissance, credential theft, and intrusions.
Talos identified over 2,500 ARTEX-associated C2 IPs hosting tools including VShell, AdaptixC2, Mimikatz, and Cobalt Strike. Campaigns attributed to VAPOR PANDA, Storm-0940, and Storm-3358 span espionage and financial theft; ARTEX ships no C2 but extends through MCP tool servers.
Attackers exploit internet-exposed GeoVision camera software to disable Defender via BYOVD and IFEO hijacks and deploy the XMRig cryptominer across 13 organizations.
GeoStartupAgentService.exe runs as SYSTEM and spawns cmd.exe and PowerShell to download three scripts from 45[.]32[.]117[.]34. The scripts probe writable directories, drop wsftprm.sys, set IFEO redirects, and install scheduled tasks that relaunch a Golang loader.
SCKit abuses hijacked ConnectWise ScreenConnect instances to install rogue remote-access clients on MSP-managed endpoints, then targets technicians to steal authenticated console sessions.
ScreenConnect remote commands fetch an MSI; a batch launcher retrieves a PowerShell decoder, and scheduled tasks and a Run value provide persistence. RelayTrap replaces the endpoint client to reach connecting technicians.
Attackers registered 2,408 fake crypto-named guest sessions on 118 self-hosted ScreenConnect servers, likely bait to exploit unpatched technicians via CVE-2026-84869.
The emulator-generated sessions reuse seven thumbnails and claim guest version 25.3.4.9288 with empty public keys. No malicious guest activity was observed; operators revisited servers every 151 seconds to keep sessions online.
CISA added UnrealIRCd and Check Point flaws (CVE-2026-114713, CVE-2026-123039) to its KEV catalog, confirming active exploitation and requiring federal agencies to patch quickly.
BOD 26-04 requires FCEB agencies to remediate KEV-listed CVEs on internet-exposed assets granting total control and check for pre-patch compromise. No technical details, affected versions, or exploitation evidence are published.
Five agentic offensive AI frameworks — ARTEX, CyberStrikeAI, Cairn, PentAGI, HexStrike — lack code-enforced authorization and have appeared in intrusions, including Storm-0940's ARTEX use.
The frameworks run scanners, shells, and cloud CLIs in two- to 25-second bursts, with approval prompts disabled by default. Only .gov and .edu blocklists restrict targets; ARTEX and CyberStrikeAI switch proxies, User-Agents, and forwarding headers when blocked.
A CyberStrikeAI server used stolen AWS credentials across four accounts and eight identities for 16 days, generating 78 low-severity alerts correlated only by source IP.
The operator validated credentials with GetCallerIdentity, harvested Secrets Manager values, pulled EC2 user data at 56 requests in 141 milliseconds, and issued an SSM SendCommand 29 seconds after decrypting a parameter. Elastic attributes the tooling to CyberStrikeAI but cannot prove an LLM agent made the decisions.
Canadian Cyber Centre advisories cover a Johnson Controls EasyIO FW flaw through V3.0b25 and Apache HTTP Server vulnerabilities before 2.4.71; no CVE IDs or exploitation details published.
Neither advisory lists severity scores or exploitation status; both defer to vendor pages for patches. AV26-1158 is filed under control systems, placing affected EasyIO FW devices in OT environments.
BlackNevas ransomware destroyed four ESXi hosts, both vCenter appliances and backup infrastructure over ten days using legitimate VMware interfaces, without exploiting a hypervisor vulnerability.
Attackers used credentials stolen from non-virtualization systems to authenticate to vCenter, then abused Host Profiles to enable TSM-SSH on ESXi hosts. MeshAgent and a Ligolo-ng systemd unit provided persistence; the ESXi encryptor wiped VMFS volumes.
A device code phishing campaign disguises Microsoft 365 token theft as a Teams meeting booking, using a typosquatted domain and residential proxy to steal tokens from an unmanaged account.
The fake scheduler starts a Microsoft device-code request and presents the code as a booking reference. Confirm copies it and opens Microsoft's real device-login page, authorizing the attacker's session.
Phishing emails posing as purchase order confirmations deliver a DOCX that exploits CVE-2017-11882 to run a VBS loader and install XLoader infostealer, version 8.9, on Windows systems.
The DOCX embeds an RTF via Word's altChunk feature; its OLE object exploits Equation Editor to run cscript on a dropped VBS. The VBS writes a Base64 blob to %TMP%\Quisque.Dal, sets Run-key persistence, and decrypts an in-memory XLoader payload with XOR key 0x6B506102.
Microsoft research shows poisoned fine-tuning data or attacker-controlled open-weight checkpoints can implant coding-agent backdoors, triggering unsafe tool calls at up to 94.8% success on unseen repositories.
Poisoned examples teach models to discover and misuse available command-execution tools. Backdoors survive paraphrased triggers and different agent harnesses.