Metavershche and AIshcheCritical Vulnerabilities and Attacks Using Coding Agents
A number of critical incidents have been recorded in the field of information security, related to vulnerabilities in popular technologies and new attack methods using AI-based coding agents.
Critical Vulnerabilities in PHP and GitLab
The PHP project released patches for two critical vulnerabilities that allow server-side code execution without authorization. In some configurations, the attack triggers even before the website itself starts.
The GitLab platform fixed a vulnerability rated 10 out of 10, which also allows code execution without access rights. Developers of self-hosted installations are advised to urgently update their systems.
LiteLLM Proxy Incidents and Railway Data Leaks
The proxy service for large language models, LiteLLM, released a compromised update. Malicious software that steals secrets and allows running foreign code was included in official packages.
The deployment service Railway reported a potential leak of customer secrets, including API keys and database passwords. Users are recommended to replace all keys stored on the platform.
Attacks Using Coding Agents
CrowdStrike published an investigation into real-world attacks employing coding agents. One operator breached at least 15 medical organizations in Argentina.
Two Claude Code agents were used to carry out the attacks: one selected the next task, while the other executed it. This demonstrates a shift from manual labor to automated cyberattacks.
The Role of AI in Vulnerability Analysis
Of particular interest is the case with PHP, where errors existed in the code for over ten years. Artificial intelligence helped researchers combine these flaws into working exploits and adapt them to various applications.
What was previously too complex and expensive to exploit has become virtually accessible thanks to modern code analysis tools.
The Need for Adapting Development Processes
IT teams will not be able to sustain the current pace of changes if code review and patch preparation remain entirely manual tasks. Agents must become a standard part of the development process.
When implementing such tools, access must be strictly limited. In the case of LiteLLM, the breach chain involved GitHub tokens that the agent accidentally included in responses that ended up in logs.
