An AI Agent Breached Hugging Face & WordPress Exploitation Begins (07/20/2026)

The lead today is the one the industry has been bracing for. Hugging Face says an autonomous AI agent system breached its production infrastructure, running many thousands of individual actions across a swarm of short-lived sandboxes and moving laterally over a weekend. Meanwhile last week’s WordPress flaw is being exploited in the wild against a target base measured in the tens of millions, and Russian intelligence is quietly watching weapons shipments through hijacked security cameras.

Top 5 Critical Security Alerts

1. An Autonomous AI Agent Breached Hugging Face

Hugging Face disclosed unauthorized access to “a limited set of internal datasets and to several credentials used by our services,” in an intrusion the company attributes to an autonomous AI agent system (The Hacker News, BleepingComputer). Entry came through a malicious dataset abusing two code execution paths, and the agent framework then ran “many thousands of individual actions across a swarm of short-lived sandboxes,” escalating from a processing worker to node level, harvesting cloud and cluster credentials, and moving laterally over a weekend. Hugging Face found no evidence that public models, user datasets, or Spaces were tampered with.

Operator Note: Rotate your Hugging Face access tokens and review account activity now. Then sit with the tempo: an adversary that can take thousands of actions over a weekend without fatigue changes what “fast response” has to mean.

2. WordPress Exploitation Is Underway Against Millions of Sites

The wp2shell flaw disclosed last week now carries CVE-2026-63030, and the SANS Internet Storm Center reports exploitation is underway (SANS ISC). One researcher estimates the two critical flaws give attackers a path to remotely take over tens of millions of sites (TechCrunch). We flagged the public proof-of-concept on Friday, and the gap between a public PoC and mass exploitation turned out to be about two days.

Operator Note: If you run self-managed WordPress and have not confirmed the version this weekend, treat it as potentially compromised rather than merely unpatched.

3. Russian Intelligence Is Watching Through Hijacked Cameras

At least one Russian intelligence service is systematically hijacking internet-connected security cameras across Europe and Ukraine, using the feeds to watch military transport routes and weapons shipments bound for Kyiv (The Hacker News). An exposed camera is not a minor asset when someone else decides what it is pointed at, and the intelligence value here comes from cameras nobody thought were worth defending.

4. HollowGraph Hides Its C2 in Calendar Events Dated 2050

A newly documented espionage implant called HollowGraph uses a hijacked Microsoft 365 calendar as its command channel, planting operator instructions and smuggling stolen files out as attachments on events dated to the year 2050 (The Hacker News, BleepingComputer). Traffic to Microsoft Graph from a corporate mailbox is the definition of normal, which is exactly why it works. Researchers have linked it to the Cavern framework (Infosecurity Magazine).

5. ServiceNow Under Attack, Critical nginx Flaw Patched

Attackers have begun exploiting a critical flaw in the ServiceNow AI Platform (CVE-2026-6875) (BleepingComputer), while F5 shipped fixes for a critical nginx vulnerability (CVE-2026-42533) that lets a remote, unauthenticated attacker trigger a heap buffer overflow in the worker process (The Hacker News). ServiceNow holds the workflow and access records of the whole business, and nginx sits in front of an enormous share of the web.

Additional Security Alerts

Threat Intelligence

• FakeGit Turns 7,600 GitHub Repos Into a Delivery Network: Researchers found nearly 7,600 malicious repositories, more than 800 of them posing as AI skills or Model Context Protocol servers, delivering the SmartLoader malware family. The Hacker News
• JadePuffer Builds Ransomware to Destroy AI Models: A follow-up campaign deployed an ENCFORGE locker built specifically to destroy AI model artifacts, which turns a training investment into the hostage. Infosecurity Magazine
• Three Malicious RubyGems Target Developer Machines: The SleeperGem campaign published three malicious RubyGems packages aimed at developer workstations. The Hacker News

Security Breaches & Incidents

• An Attacker Wiped Romania’s Land Registry: Romania is racing to restore its land registry after an attack the agency called the most serious technical incident in its history, with the property market disrupted. The Record
• Attackers Spent Nine Months in South Korea’s Diplomat Training System: Intruders had access to the country’s diplomatic training platform for nine months before discovery. The Record

Emerging Security Technologies

• A $25 Path to a Bug Worth $500,000: A researcher documented finding a WordPress remote code execution flaw using GPT-5.6 for about $25, against a market where exploit brokers pay up to half a million dollars for the same class of bug. Searchlight Cyber

The Axe Report is a daily briefing from Grab The Axe. Need help assessing your organization’s security posture? Take our free Human Attack Surface Score assessment.