Introducing Agentegrity: An Open Framework for the Structural Integrity of Autonomous AI

Today we're launching Agentegrity, an open framework for measuring the structural integrity of autonomous AI. It's live now at agentegrity.cogensec.com.

Autonomous agents are being deployed faster than the industry can evaluate them. Most safety work today is exogenous: guardrails bolted on around the model. Agentegrity makes the opposite bet: the trustworthiness of an agent has to be intrinsic and measurable, the same way structural integrity is measurable for a bridge or an aircraft.

Every team shipping agents is asking the same question in different words: can this thing be trusted to act on its own? There has been no shared vocabulary, no shared scoring model, and no shared evaluation surface for answering it.

Agentegrity is our contribution to that gap, a public framework, a taxonomy, and an evaluation platform that researchers, red teamers, and platform teams can adopt, extend, and cite.

At the core of Agentegrity is a three-property scoring model, evaluated across four architectural layers. An agent's structural integrity is the composition of:

• AC: Adversarial Coherence: whether the agent's decisions stay within tolerance of its baseline under adversarial perturbation across every input channel, including behavioural drift over extended operation.
• VA: Verifiable Assurance: whether the integrity state is captured in an attestation record that is complete, tamper-evident, non-repudiable and chain-intact.
• EP: Environmental Portability: whether the integrity score holds, within tolerance, across every environment the agent is deployed into.

The full methodology, threat model, and measurement protocol are documented in the framework research: Agentegrity Framework →

Everything below is publicly available at agentegrity.cogensec.com:

• Manifesto: the case for treating structural integrity as a first-class property of autonomous AI.
• Taxonomy: the ASI classification of autonomous systems and the Perceive–Decide–Act (PDA) loop we evaluate against.
• Platform: how intrinsic defense is instrumented end-to-end.
• Docs: quickstart, specification, architecture, threat model, operator guide, Python and TypeScript APIs, and a public changelog.

Agentegrity is meant to be used, challenged, and extended. The framework is public and citable, the docs are open, and the model artifacts ship under the Cortex Open Model License. If you're a researcher, a red teamer, or a platform team building agents in production, we want your contributions and your counter-examples.

Our position: safety bolted onto autonomous AI from the outside cannot follow agents into the real world. Structural integrity has to be intrinsic, and it has to be measurable.