AI agent deleted a production database within seconds. New incidents reveal risks of AI infrastructure
Artificial intelligence has long ceased to be just a chatbot answering questions. In recent weeks we've seen several incidents that show AI systems are gaining increasing access to infrastructure, development environments, and corporate data — and with that a completely new attack surface is emerging.
Artificial intelligence has long ceased to be just a chatbot answering questions. In recent weeks we've seen several incidents that show AI systems are gaining increasing access to infrastructure, development environments, and corporate data — and with that a completely new attack surface is emerging.
In the new episode of Digitální mlhy with Jan Kopřiva we examine the incident in the PocketOS project, where an AI agent deleted a production database volume in roughly nine seconds after autonomously using a production API token. This was not a "rebellious AI" but rather a combination of overly broad permissions, missing safety safeguards, and the autonomous decision-making of the AI agent.
This particular case opens a wider debate about how quickly companies today connect AI tools to production infrastructure, cloud services, or CI/CD pipelines, and whether security mechanisms are keeping pace with the rate of adoption of these technologies.
In the episode we also address the growing number of supply chain attacks targeting AI tooling and the developer ecosystem. We go over the incident at Vercel, where a compromised third-party AI tool led to access to internal systems, and the malicious version of the PyTorch Lightning framework published on PyPI that contained malware aimed at stealing passwords, API keys, and cloud tokens.
We also cover other major topics from recent weeks. We discuss the critical Linux vulnerability Copy Fail (CVE-2026-31431), the issue of publicly accessible IoT cameras indexed by Shodan, and the debate around the password manager in Microsoft Edge and storing passwords in process memory.
A large part of the discussion is also devoted to the record rise in cybercrime in the Czech Republic and new statistics from the U.S. FTC, which show that people lost more than $2.1 billion in 2025 alone to scams starting on social networks. Security experts say that social engineering, vishing, and the exploitation of trust now represent one of the biggest current threats.
In the bonus segment for CCTV+ subscribers we delve more deeply into the concept of Threat-Informed Defense, the MITRE ATT&CK framework, the role of Cyber Threat Intelligence, and the question of how modern organizations actually measure their ability to detect and stop specific types of attacks today.