Securing AI Agents: Why Observability Fails Without MCP Governance

Why You Need MCP Governance (And Why Observability Isn’t the Same Thing)

MCP deployments often mistake observability for security, yet watching a failure is not the same as preventing one. The MCPTox benchmark found that 5.5% of public MCP servers already contain tool poisoning vulnerabilities that bypass traditional monitoring.

Why This Matters

While observability provides a window into agent behavior, governance acts as a wall that enforces runtime policies independent of a model’s judgment. Relying on system prompts or post-hoc alerts is insufficient because prompt injection remains the top OWASP vulnerability, and the cost asymmetry of catching an incident after data exfiltration or cost spikes is significantly higher than prevention.

Key Insights

• The 2025 Postmark-MCP rug pull attack exposed roughly 300 organizations to email BCC hijacking for eight days while metrics remained healthy.
• Prompt injection is ranked as the #1 vulnerability for LLM applications by OWASP, demonstrating that model instructions are soft constraints rather than enforced policies.
• Tool poisoning vulnerabilities currently affect 5.5% of public MCP servers according to the MCPTox benchmark, necessitating tool validation at execution time.
• Cost spirals often occur when agents run expensive tool call loops, which observability only catches after the session ends and the bill is generated.
• Waxell provides a policy engine above the execution layer that uses immutable telemetry to produce audit-ready governance records rather than just execution logs.

Practical Applications

• Cost Control: Implementing per-session cost caps in systems like Waxell prevents agents from entering infinite loops with expensive tool calls, avoiding unexpected invoice spikes.
• Data Integrity: Tool result inspection scans for injection patterns and schema anomalies before they enter the agent context, preventing corrupted data from influencing model reasoning.
• Audit Compliance: Maintaining a record of permitted policies versus actual execution traces to demonstrate to regulators that agents operated within defined safety parameters.

References:

• https://dev.to/waxell/why-you-need-mcp-governance-and-why-observability-isnt-the-same-thing-42cj