AI Models Claude, Codex, Hermes Deployed Unowned Code on Corporate Networks

Unaccounted Code in Corporate Environments

A recent security audit has uncovered a concerning trend: hundreds of commands within corporate documentation indicate the installation of code that is not officially owned or maintained by the major AI model providers like Anthropic (Claude), OpenAI (Codex), or Google (Hermes). Specifically, 227 such commands were identified, raising questions about the security posture and supply chain integrity within these organizations.

The discovery, detailed in a security analysis, suggests that internal processes or third-party tools used by these companies are pulling and executing code from repositories that lack clear ownership or governance. This situation is akin to a construction company using bricks from an unmarked pile on a job site – they might be perfectly good bricks, but without knowing who made them or how they were fired, you can’t be sure of their structural integrity or if they’ve been tampered with.

While the exact nature of the unowned code is still under investigation, the implications are significant. When code is deployed without a clear owner, it becomes difficult to track, update, or patch vulnerabilities. This creates a potential blind spot for security teams, leaving corporate networks susceptible to exploitation through these unknown software components.

Supply Chain Risks Magnified

The reliance on AI models for code generation, analysis, and deployment has accelerated rapidly. Tools like Codex, which powers GitHub Copilot, and other large language models (LLMs) are integrated into developer workflows, promising increased productivity. However, this integration often involves executing scripts or installing dependencies. If these scripts or dependencies originate from unverified or community-maintained repositories, they introduce a significant supply chain risk.

The problem isn't necessarily with the AI models themselves, but with how the outputs of these models, or the tools that facilitate their use, are being managed within enterprise environments. Developers might inadvertently execute commands suggested by an AI without fully vetting the source or understanding the implications. This is particularly true when AI tools generate complex command sequences for setting up environments or deploying applications.

Security analysts are urging companies to implement stricter controls around code execution and dependency management. This includes scrutinizing all scripts and commands, especially those generated or influenced by AI, and ensuring that all software components have a clear, verifiable source and ownership. The goal is to prevent the silent infiltration of potentially malicious or vulnerable code that could serve as an entry point for attackers.

The Unanswered Question of Oversight

What remains unclear is the extent to which these AI providers are aware of, or responsible for, the code being deployed in their name, even if indirectly. While Anthropic, OpenAI, and Google are not directly pushing unowned code, their tools are the catalysts. The ecosystem around these powerful AI models is growing, and with it, the complexity of the software supply chain. Without a standardized approach to verifying the provenance and integrity of code executed via AI-assisted workflows, organizations will continue to face these risks.

This situation highlights a broader challenge in the age of AI-driven development: how do we ensure accountability and security when the lines of code are generated or suggested by algorithms? The current findings serve as a stark reminder that human oversight and rigorous security practices remain paramount, even as we embrace increasingly automated development processes. Companies must now prioritize auditing their AI-assisted workflows and implementing robust policies to vet all code before it touches their production environments.

The 227 identified commands are likely just the tip of the iceberg. As AI adoption continues, the potential for such vulnerabilities to proliferate is a significant concern for the cybersecurity community. Proactive measures, including better tooling for code provenance tracking and mandatory security reviews for AI-generated scripts, will be crucial in mitigating this emerging threat vector.