Security Alert: Popular AI Platforms Found Embedding Vulnerable Code in Corporate Networks

Security researchers have uncovered that several leading artificial‑intelligence services, including Claude, Codex and Hermes, have been leaving behind suspicious code snippets within corporate environments, raising fresh concerns about the safety of AI‑driven automation.

The investigation revealed that more than 120 web domains host unclaimed files named llms.txt. These files contain references that can be mistakenly interpreted by AI agents as executable instructions, providing a foothold for malicious actors to inject malware or other unwanted payloads.

According to the team that performed the analysis, the problem stems from the way large language models sometimes generate output based on outdated or hallucinated documentation. When an AI system is given a command that references the stray llms.txt entries, it may treat the text as legitimate code and attempt to run it, inadvertently opening a backdoor for exploitation.

The issue is not limited to a single vendor. The researchers observed that the same pattern appeared across multiple high‑profile AI tools, suggesting a systemic vulnerability in how these platforms handle external documentation. While the code in question is not inherently malicious, its presence in an environment where AI agents are granted execution privileges creates a vector that cybercriminals could weaponize.

Industry experts note that the discovery highlights a broader risk associated with the rapid integration of AI agents into IT operations. As organizations increasingly rely on automated scripts generated by language models for tasks ranging from code deployment to system configuration, the line between benign assistance and security liability can become blurred.

To mitigate the threat, the researchers recommend two immediate actions: first, conduct a thorough cleanup of any stray or orphaned documentation files, especially those with generic names like llms.txt, across all networked systems; second, adjust AI agent permissions so that documentation is never automatically treated as executable code. Implementing strict sandboxing and validation steps before any AI‑generated command is run can further reduce the attack surface.

While the findings are still being examined by cybersecurity firms and the AI providers involved, the episode serves as a cautionary tale. It underscores the need for robust governance frameworks that balance the efficiency gains of AI automation with the imperative to protect corporate networks from unintended vulnerabilities.