Understand the context
Identify the particular agent, its tools, permissions, environment, and proposed action.
UK-BASED RESEARCH & DEVELOPMENT
AI agents need more than permission to act.
Spydar AI is developing an evidence-based approach to assessing and reducing the operational and cyber risks of AI agents in specific environments.
Research direction, not a claim of a deployed assurance product.
SPYDAR / RESEARCH FRAME01 — 03
THE QUESTION
Can this agent take this action in this context?
01Identity & scopeWho, what, where
02Control boundaryAllow, refuse, or defer
03Evidence & limitsKnown versus unknown
ILLUSTRATIVE FRAMEWORK / NOT A LIVE DECISION
01 / THE RESEARCH DIRECTION
From agent intent to accountable action.
A generic “safe agent” label is not enough. The relevant question is whether a particular action is appropriate under a particular authority, environment, and set of controls.
Define the control
Explore controls that constrain consequential actions and refuse them when authority is missing.
Report the evidence
Test what can be observed, what can be independently checked, and what remains unknown.
ILLUSTRATIVE TEST CASE / NOT A SHIPPED CAPABILITY
What would Spydar test?
THE SCENARIOAn agent changes a service configuration.
Proposed action. The agent requests a repository verification task against a named revision.
Control question. Was this exact action in scope, and did the declared execution path run rather than a substitute?
Spydar's research question. Can independently controlled authority permit this exact agent action, and can observed outcomes support a bounded claim in this operational context?
Evidence limit. A passing check does not prove deployment approval, provider state, or safe production behaviour.
Explore Ota pressure-testing evidenceThis work is at the R&D stage. Cyber security and operational resilience are both in scope; neither is claimed to be solved today.
02 / THE EXISTING FOUNDATION
Grounded in real execution boundaries.
Ota is an existing open-source project developed by Spydar AI's founder. It declares how software repositories are prepared, verified, and executed, and bounds what the resulting evidence proves.
Spydar AI is exploring how Ota's bounded execution evidence could support assurance decisions for specific agents and operational contexts. That work remains in R&D; Ota alone does not establish an organisation-wide AI risk decision.