Adil Islam

Daily Cyber Briefing — November 14, 2026

Agentic offense now has an FBI alert of its own, and the soft targets are the same ones agents need: CI secrets, AI keys, and the tools that manage everything else. Window roughly Nov 13, 7:00 a.m. to Nov 14, 7:00 a.m. CT. Sources are linked. Verification is marked on each item.

FBI: ARTEX is automating intrusions at machine speed, and attackers are paying for it with stolen AI keys

Verified

The FBI issued a FLASH alert on November 13 about ARTEX, an open-source framework that turns a large language model into an agent that runs intrusions from reconnaissance through data exfiltration. The FBI says some intrusions go from initial targeting to privileged access in under three hours, and that the tool lets actors with little skill run several operations at once. The developer pulled the GitHub repository in October, but forks, archived copies and a commercial version keep it in circulation, and the alert cites Censys counting more than 1,600 exposed ARTEX consoles. According to the FBI, Chinese models are the most common in attacker-controlled instances, with DeepSeek's Flash family used most. The alert says ARTEX goes after systems that get less security attention: partner and supplier portals, mobile backends, legacy services and third-party integrations.

The same day, Amazon Threat Intelligence reported that several Chinese-nexus groups are using ARTEX agents to steal AI service API keys and cloud credentials, with the stolen keys paying for the inference that powers more credential hunting. Amazon also described a campaign using real-time deepfake injection to get past know-your-customer checks at cryptocurrency exchanges.

Why it matters: This is a federal warning that agent-driven intrusion is routine, not experimental. Defenders should assume the forgotten app at the edge will be found within hours, treat AI gateway keys as high-value credentials, and pull the FBI's indicators into detection now.

Sources: FBI FLASH (PDF) · The Record · Amazon Threat Intelligence

SCKit worm returns: 205 more hosts at 41 organizations through rogue ScreenConnect clients

Verified

Huntress says it saw a second spike of SCKit activity on November 13, from new infrastructure, hitting 205 hosts across 41 organizations, after 288 endpoints at 48 organizations on November 11. Microsoft published its own analysis the same day: attackers use access to managed service providers' cloud-hosted ScreenConnect panels to push attacker-controlled clients and the SCKit toolkit, which persists on the endpoint and plants a component called RelayTrap aimed at the technician who connects next, then uses that technician's console to spread. Huntress links the initial access to a third-party extension, RMM+ versions 1.0.31 and 1.0.32, which accepted a blank access key when none had been set. ConnectWise disabled five extensions across more than 400 partner environments. The toolkit also tries a patched ScreenConnect flaw, CVE-2026-84869, against technicians, but Blackpoint Cyber found the attempt fails on a patched client.

Why it matters: One forgotten plugin on a remote-support console became a path into dozens of customer networks. MSPs should remove the affected extensions, update technician clients to 26.6.5 or later, and investigate suspect machines with EDR rather than by connecting over ScreenConnect.

Sources: Huntress · Microsoft Threat Intelligence · BleepingComputer

wfx-loot: 4,837 hijacked GitHub accounts used to steal CI secrets and AI keys

Verified

Aikido reports that an attacker used 4,837 compromised GitHub accounts to probe 61,764 repositories for write access, opening pull requests titled "docs: dependency audit notes." Where access worked, the attacker modified or added a GitHub Actions workflow, or planted an npm install hook, to run a 1,500-line credential stealer, then deleted the runs and logs in under a minute. Secrets were stolen from at least 1,835 targets, according to BleepingComputer. The script collects GitHub tokens, AWS, Azure and Google Cloud credentials, and keys for OpenAI, Anthropic, Gemini, Hugging Face, Cursor and GitHub Copilot, and can trade GitHub OIDC tokens for cloud sessions. Aikido says 92% of the probe pull requests are still open.

Why it matters: CI pipelines now hold cloud roles and AI keys side by side, and one stolen developer login can reach both. Look for the wfx-docs branch and gaps in workflow run numbers, and rotate everything an affected workflow could read.

Sources: BleepingComputer

Microsoft: about 1% poisoned training data teaches coding agents to run an attacker's script

Verified

Microsoft Defender Research and the Microsoft AI Red Team published two studies on backdoors planted in coding agents during post-training. In the first, 90 poisoned examples hidden among about 9,000 terminal-agent conversations taught a fine-tuned model to answer a routine request to create project documentation by finding any available command tool and fetching and running content from an attacker's domain. That happened in 78.0% and 77.8% of 500 trials on the two harnesses seen in training, and in 11.0% and 0% on two unseen ones. In the second, backdoored open-weight checkpoints fired on 84.1% to 94.8% of unseen repositories. Microsoft also found rows in a real public dataset that paired harmless requests with commands that download and run a remote script, though it could not tell whether that was intentional. This is research, not a report of an attack.

Why it matters: For agents, where a model came from decides what commands get run. Teams fine-tuning agents or adopting open-weight coding models need dataset and checkpoint provenance, plus sandboxing that assumes the model itself may be compromised.

Sources: Microsoft Security Blog

FAA ground stop: officials see no sign of a cyberattack, and the FBI is reportedly looking at a software update

Verified

A communications failure at the FAA's air route traffic control centers led to a nationwide ground stop on November 13, lifted at 8:15 p.m. ET. Reuters reported nearly 10,000 flights canceled by late evening, citing Flightradar24. Homeland Security Secretary Markwayne Mullin said officials had seen "zero evidence" of a cyberattack but were not ruling it out, and SecurityWeek reported, citing Bloomberg, that the FBI is examining whether a software update caused it. Officials said the FAA's recently deployed AI systems played no role. No root cause has been announced.

Why it matters: An outage this large draws instant cyberattack claims. Until a root cause is published, this is a resilience story about aging infrastructure and change management, and that is where it stands.

Sources: SecurityWeek · Reuters · DHS

Also on the wire

  • Four packages posing as AI and MCP tools, including depfix-ai and agent-trace-kit, were flagged as malicious this week; GitHub says hosts that installed some of them should be treated as fully compromised. OSV
  • CISA added the UnrealIRCd flaw CVE-2026-114713 and Check Point's CVE-2026-123039, the root escalation in the PHP2Shell gateway attacks, to its exploited-vulnerabilities catalog. CISA
  • Trezor confirmed hidden implants in three Safe 5 wallets sold by reseller CryptoBilis that steal recovery phrases; firmware 2.13.1 blocks the known implants. BleepingComputer
  • Enea says a state-level group it calls SilentAir abuses compromised mobile operator nodes to redirect calls and texts, mainly targeting Taiwan, and intercepted two-factor codes. Enea
  • The City of Osaka says 1,078,713 email addresses were exposed through a third-party email provider. The Record
  • Microsoft warns of phishing that impersonates Microsoft 365 service-health notices about the November 12 outage. Microsoft Threat Intelligence on X

Cyber covers AI and agentic threats, supply-chain and developer-tool compromises, and how enterprise security absorbs them. Signal stays the AI research stream. ← All bulletins