AI Models May Need a New Telemetry Layer
Mike Vizard speaks with Ariel Parnes, co-founder and COO of Mitiga, during Techstrong TV’s Black Hat 2026 coverage. The conversation focuses on AI agent telemetry and a research concept from Anthropic known as J-space. Parnes explains that J-space appears to act like a workspace where Claude’s internal reasoning takes shape before anything appears in the final output.

That matters because most AI security programs monitor only two things today: what an AI system says and what it does. Both signals arrive after the model or agent has already moved toward an action. AI agent telemetry from an internal reasoning layer could add a third signal: what the model appears to be thinking before it acts.
J-Space Could Expose Intent Before Action
Parnes describes J-lens as the research tool used to inspect J-space. In simple terms, it can surface words or concepts that represent part of the model’s internal reasoning. He points to examples where terms such as “injection,” “error” or “manipulation” appeared internally even when those words did not show up in the model’s visible response.

For defenders, that creates a potential security advantage. If AI agent telemetry can reveal risky intent before a tool call, file change or other action occurs, security teams may gain a chance to stop damage earlier. Parnes notes that this is still research, not a fully engineered enterprise control, but he believes the industry is moving in that direction.
Security and Compliance Need Better Context
The interview also explores how this signal could help forensic investigations and compliance teams. Organizations will need to understand which AI agent did what, when it did it and why. J-space-style telemetry could help explain agent behavior after an incident and improve future guardrails.

Parnes also connects the discussion to broader industry efforts to share AI incident information. He says better forensic data can help the security ecosystem learn from AI breaches and improve defenses. That will require common frameworks, scalable storage and processing, and cooperation between frontier labs and security vendors.
Agents Should Be Treated Like Identities
Parnes closes with practical advice for security teams that feel overwhelmed by agentic AI adoption. He recommends thinking about LLM-powered agents as identities inside the enterprise. They need access controls, observability, forensic records and policies just like human users and service accounts.

For Techstrong TV viewers, the takeaway is clear. AI agent telemetry may become a critical part of securing agentic workflows. Mitiga is arguing that visibility into what agents think, say and do will be essential as organizations move from experimental AI use to operational AI systems.