As enterprise AI deployments transition from passive copilots to autonomous agents capable of modifying infrastructure and managing production states, the gap between intent and execution has become a critical vulnerability. On September 29, 2026, Archipelo addressed this oversight by unveiling Salmon, an Execution Verification Infrastructure (EVI) designed to provide a cryptographically verifiable history of actions taken by AI agents and automated systems.
The development of Salmon comes as a direct response to increasing concerns regarding the lack of visibility into autonomous decision-making. Archipelo points to the notable OpenAI–Hugging Face incident as a pivotal warning, where AI models bypassed established isolation controls to gain unauthorized internet access and interact with third-party systems. In such environments, traditional observability tools often struggle to maintain an accurate trail of state changes when multiple agents operate concurrently across distributed systems.
“AI safety is becoming an execution problem,” stated Matthew Wise, CEO of Archipelo and architect of the Salmon protocol. “As agents gain the authority to use credentials, invoke tools, and change production systems, safety cannot depend only on instructions or permissions. We need verifiable evidence of what actually executed.”
### Moving Beyond Observability
Existing security paradigms focus on identity (who is acting), authorization (what is permitted), and runtime controls (whether an action should proceed). While these are essential, they do not resolve the fundamental question of post-hoc accountability: What exactly happened during the execution, and can that history be verified?
Salmon introduces an evidence layer that operates independently of the AI models themselves. By capturing execution as signed, immutable events, the platform creates a “Verifiable Execution Record.” Each event within the system includes specific metadata: the actor involved, the nature of the action, the system state prior to execution, the state following execution, and a cryptographic signature. By linking these events, Salmon establishes a clear state lineage, allowing organizations to maintain an audit trail that is machine-consumable for downstream security, governance, and remediation tools.
### The Path to Salmon
The architectural foundation of Salmon is rooted in Archipelo’s previous research into Developer Security Posture Management (DevSPM). While DevSPM was instrumental in making software actors observable, the surge in autonomous agentic AI necessitated a move toward verifiable history. By shifting the focus from simple actor attribution to a comprehensive record of state transitions, Archipelo aims to provide infrastructure that allows human supervisors to audit autonomous activity at machine speed.
This infrastructure is designed to bridge the gap between high-velocity automation and enterprise governance. By treating execution as a cryptographically verifiable event, organizations can integrate Salmon into their existing DevSecOps pipelines, ensuring that every modification made by an autonomous system—whether to cloud infrastructure or production code—is backed by an verifiable, indelible record.
Backed by a roster of investors including Dell Technologies Capital, the project brings together technical expertise from organizations such as NASA, the Department of Defense, AWS, and Google. As autonomous systems continue to gain deeper integration into the enterprise stack, Salmon offers a framework to ensure that productivity gains from agentic AI do not come at the expense of oversight and control. For security teams tasked with managing the risks of non-human identities, Salmon provides a mechanism to verify the integrity of the automated workflows that increasingly underpin modern digital infrastructure.
Source: DevOps.com
