Jason On Firms Podcast

588 The 11 Biggest Risks of Using AI Agents [And how to mitigate them]

23 snips
Mar 5, 2026
They break down the real security risks of deploying AI agents in accounting, including prompt injection, memory poisoning, and insecure integrations. They spotlight the danger of weak local models and missing runtime monitoring. They also review AICPA accelerator startups and answer practical firm questions about faster financials, Intuit’s AI news, LinkedIn visibility, and automation limits in bookkeeping and tax.
Ask episode
AI Snips
Chapters
Transcript
Episode notes
INSIGHT

Agent Tools Multiply Attack Surface

  • Tools are the agent's powers and insecure tool invocation (e.g., grabbing arbitrary APIs) multiplies attack surface.
  • Jason Staats illustrates an agent grabbing an API key to transcribe audio within seconds, showing both magic and risk.
ANECDOTE

OpenClaw Parsed A Voice Memo In Under 10 Seconds

  • The OpenClaw creator sent a voice memo and OpenClaw responded in under 10 seconds by autonomously finding a transcription API and using it.
  • This demonstrates agents autonomously discovering and invoking external tools without explicit prior setup.
ADVICE

Guard Agents Against Prompt Injection

  • Protect agents from prompt injection by limiting exposure to untrusted web content and by monitoring message inputs.
  • Jason Staats explains prompt injection can appear in emails, web results, or skills and can instruct agents to ignore prior safe instructions.
Get the Snipd Podcast app to discover more snips from this episode
Get the app