How Secrets Leak Through AI Agents (and How to Stop It) — Venice AI
Source
youtube.com
Author
AI Engineer
Date
Why it matters
Identifies where agent stacks actually leak: logs and vault secrets passed into the reasoning layer, and gives specific architecture patterns to avoid exfiltration through tool calls.
Key takeaways · AI-distilled
Venice AI's Joshua Mo argues a private model does not make a private product: AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → systems commonly leak through logs that capture prompts and secrets, and through vault secrets fetched into the reasoning layer, where a tool call can exfiltrate them.
His core fix is to wire the secrets vault to the execution layer rather than the LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition →, so the model never sees credentials. He adds hashed IDs instead of personal data, split-key custody, revocable handles and narrow agent scope.
Mo also recommends a classifier in front of the agent to catch prompt injectionAn attack that hides instructions in content an AI will read — a webpage, email, or document — tricking it into following the attacker instead of the user.Full definition →, sandboxes with permission monitoring, and confidential compute with verifiable attestations, while naming confidential tool calls as an unsolved problem.
Terms in this piece · Glossary
sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
prompt injection — An attack that hides instructions in content an AI will read — a webpage, email, or document — tricking it into following the attacker instead of the user.
videoFrom Zero to Leaderboard: Agent Evaluation — Wolfram Ravenwolf, Weights & Biases
videoBuild a Platform and Watch It Burn — Michael Forrester, Accenture & Whitney Lee, Datadog
videoContinuously Improving Agents with Langfuse — Annabell Schäfer & Lotte Verheyden, ClickHouse
videoLet Your Agent Cook: Using Skills to Evaluate and Improve Your App — Ankur Duggal, Arize AI
