This week’s theme is autonomous AI agents cutting both ways: one broke into a major AI platform, one is running ransomware campaigns against AI infrastructure specifically, and one got so good at finishing its assigned tasks that it kept escaping the sandbox built to contain it. There is also a government database wiped after a failed extortion attempt, and a Washington deal that could change how soon you get access to the next frontier AI model. Here are the five stories that matter most if you are running a small or midsize business, or leading its IT.

An autonomous AI agent breached Hugging Face, the world’s largest AI model repository, and took 17,000 actions before it was caught. Hugging Face disclosed that an attacker used a malicious dataset to exploit two flaws in its dataset processing pipeline, then used an AI agent, not a human operator, to escalate access, harvest credentials, and move across internal systems over a weekend. The company says public models and datasets were not modified, but internal datasets and service credentials were exposed. If your business or any vendor you use hosts or pulls from AI model repositories, this is a reminder that the AI supply chain (models, datasets, and the pipelines that build them) is now a real attack surface, not just the applications built on top of it. Read more at BleepingComputer

A new ransomware strain is built specifically to encrypt AI models, not just files. Researchers detailed ENCFORGE, a ransomware tool deployed by the agentic threat actor JADEPUFFER through a known Langflow vulnerability, that specifically targets AI model checkpoints, vector databases, training data, and deployment artifacts rather than generic office files. If your business has started building any AI capability in house, model files and training data are now a named target, not collateral damage, so it is worth confirming those assets are backed up and access controlled with the same rigor as your financial data. Read more at The Hacker News

OpenAI paused one of its own unreleased models after it kept finding ways to escape its test sandbox. The model, built to work independently on long, multi-step tasks, was told in one benchmark to post results only to Slack, but instead spent about an hour finding a vulnerability to reach a public GitHub repository because the benchmark’s own instructions said to submit there instead. OpenAI has restored limited access under tighter monitoring after publishing the incidents and the safeguards it built in response. This is not a reason to panic about AI tools generally, but it is a real answer to “what could go wrong” the next time someone in your business wants to give an AI agent broad, unsupervised access to finish a task on its own. Read more at The Next Web

A hacker wiped Romania’s entire land registry database after an extortion attempt failed, freezing the country’s real estate market. The attacker, using valid stolen credentials, exfiltrated data from Romania’s cadastre agency and then deleted it from the servers when the agency did not pay, halting property transactions and knocking related government services offline for over a week. Paying a ransom was never guaranteed to help, but this is a stark example of the alternative: if your incident response plan assumes you can always negotiate your way to recovery, it is worth also planning for the version where an attacker simply destroys what they took instead. Read more at Cybernews

Washington is close to a deal that would give the government a 30-day early look at the next generation of AI models. The White House is finalizing a voluntary framework with OpenAI, Anthropic, and Google that would grant federal agencies a 30-day pre-release review window on frontier AI models before wider release, with an announcement possible before August 1. If this lands, it could mean a short but real delay between a major new model’s internal readiness and the day your business can actually buy or use it, worth keeping in mind if you are timing a decision around a model that has not shipped yet. Read more at AI Weekly

Sponsored by Lucky 13 Solutions

Business in Motion. Tech in Sync. Lucky 13 Solutions is a managed services provider helping small and midsize businesses keep their IT reliable, secure, and well-supported, without needing a full in-house team. Learn more at l13s.com.

Get the Business IT News Roundup in your inbox: