Devesh JoshiCo-founder, product
Nine years building AI platforms serving 12,000+ engineers. LLM platforms, agentic systems (MCP), and multi-model safety evaluation.
Enterprise AI security: Securing autonomous agents against prompt injection and data leaks
How to implement zero-trust execution boundaries, prompt sanitization middleware, and strict rate limits for production AI agents.
The rising threat of indirect prompt injection
As autonomous AI agents read uncurated external inputs (emails, support tickets, webhooks), indirect prompt injection becomes a critical security risk. Adversaries embed instructions designed to trick models into exfiltrating database keys or modifying user records.
Relying on system prompt instructions like 'ignore malicious commands' is ineffective against sophisticated prompt exploits. Security must be enforced in the code executing around the model.
Architecting zero-trust tool boundaries
1. Read-Only Default: Grant AI agents read-only database views by default. Require explicit human-in-the-loop authorization for destructive actions (deletions, financial transfers).
2. Input Sanitization Pipelines: Pass raw user inputs through a deterministic filter that strips hidden prompt instructions before feeding payloads to the LLM context window.
3. Schema Strictness: Reject any tool parameter call that deviates from predefined Zod or JSON schemas. Never execute arbitrary shell commands or unvalidated SQL.
Topic Focus & Target Concepts
This is the work behind our AI and automation practice — agents with real grounding, voice intake, and retrieval that answers from your records rather than the model's training data.
Agents with real tool accessRunning into this in your own stack? Twenty minutes, no deck.
Book the call