AI harnesses are just middleware with old trust

Short summary

AI harnesses—LLMs wrapped with connectors and tool-calling scaffolding—face the same trust-boundary problems as traditional middleware: components blindly trusting each other's output without validation. The novel part isn't the security gap but the payload delivery mechanism: a probabilistic text generator wired directly into tool execution. The practical fix is unglamorous: schema validation, output sanitization, and least-privilege scoping at every interface hop, treating LLM output as untrusted input everywhere.

• •AI harness trust bugs are repackaged middleware security issues, not a novel AI exploit class
• •The real gap is lack of standards for component authentication/validation in agent architectures
• •Fix: treat LLM output as untrusted input at every boundary—schema validation, sanitization, least-privilege

Generated with AI, which can make mistakes.