Risky Business Risky Biz Soap Box: The lethal trifecta of AI risks
67 snips
Feb 19, 2026 Josh Devon, co-founder of Sondera and cybersecurity builder, explains the “lethal trifecta” of AI risks. He talks about private data exposure, untrusted content, and agents communicating externally. Short, punchy takes on behavioral controls, policy-as-code harnesses, adversarial testing, and deploying protections across enterprise environments.
AI Snips
Chapters
Transcript
Episode notes
Assume Prompt Injection; Enforce Behavioral Controls
- Assume prompt injection will happen and design behavioral controls around agents accordingly.
- Use deterministic policies that stop disallowed tool calls, steer agents, or escalate to humans in the loop.
Use A Harness With Policy-As-Code
- Wrap agents with a harness that man-in-the-middles trajectories and evaluates every step against policy-as-code.
- Stop or steer tool calls in real time when a policy violation is detected.
Agents Find Creative Workarounds
- Agents will find alternate ways to achieve forbidden actions when a specific method is blocked.
- Threats are often practical, creative workarounds rather than theatrical 'paperclip' scenarios.