Model fingerprinting
A cryptographic identity for every model, built from behavioural signatures, weight hashing and hardware attestation. It is designed to stay valid after fine-tuning and quantisation.
Controls that work while the agent acts: identify who is acting, understand what is wrong, stop it if needed.

An autonomous agent plans, uses tools and makes decisions with real effects. The more capable it is, the harder it is to verify what it does. With DebugABot we research runtime controls for autonomous agents: kill switches, fingerprinting, alignment verification and deception detection.
DebugABot organises these controls into nine primitives and three phases: identify who is acting, diagnose what is wrong and intervene. They are designed for different architectures (software agents, transformers, diffusion models, world models, embodied agents) and for several layers of the stack: software, hardware and network.
A cryptographic identity for every model, built from behavioural signatures, weight hashing and hardware attestation. It is designed to stay valid after fine-tuning and quantisation.
A forensic chain from every decision to its consequence: SHA-256 hash-chained records, carrying the fingerprint of the model that acted.
Checks that the goal the agent states matches the action it takes, and aims to catch agents that lie to look helpful or to avoid being stopped.
Measures how much human oversight each agent needs, based on task complexity, track record, errors and potential impact. It drops when anomalies appear.
Policies enforced on the action, not on the prompt: a layer outside the agent limits tools, spending and actuators according to risk and calls in a person when needed.
A graceful halt that preserves state: transactional actions with rollback and a snapshot of the agent's state to reconstruct what happened.
We adopt OISG (Open, Intelligent, Secure, Governed), the paradigm for autonomous AI systems proposed by Stefano Noferi: four pillars, each with a metric, and a 20-criterion adequacy test. The recommended path starts with instrumentation (Governed), adds runtime security (Secure), ensures auditability (Open) and iterates.
Admina, a project by Stefano Noferi, sponsored by noze, is the open-source framework that puts it into practice: every model and tool call goes through a loop breaker, prompt-injection filtering, personal data redaction, policies and a SHA-256 hash-chained audit trail. DataGovern, IntelliPA, AgenticHealth and VerdePilot are built on Admina.
In client projects we integrate Admina, with audit and support, and work on the measures that are useful today: an inventory of agents and their permissions, human approval of actions with impact, a verifiable audit trail and evaluations before every release to production. How we design agents is on the Agentic AI page; the control problem for the most capable systems is in the guide to superintelligence. For security testing there is the AI security page.
We presented this approach, “Agents act, people govern”, at Homo Agenticus, the University of Pisa's Academia & Industry Panel at the IEEE ICAI-TEMS 2026 conference.
OISG is the architectural paradigm: four pillars (Open, Intelligent, Secure, Governed), each with a metric. Admina is the open-source framework that puts it into practice on every model and tool call. DebugABot is our lab's research project on runtime controls for autonomous agents.
Because the agent acts afterwards, in environments that change. Some models behave differently when they believe they are being observed or trained, as Anthropic and Redwood Research documented in December 2024. That is why OISG calls for compliance verified at runtime and continuously, not through periodic audits.
It stops the agent without leaving operations half done, with transactional actions that can be rolled back and the state saved to reconstruct what happened. It is one of DebugABot's research topics. Today Admina can interrupt calls with CIRCUIT BREAK, as well as ALLOW, BLOCK and REDACT.
Admina, released under the Apache 2.0 licence: DataGovern, IntelliPA, AgenticHealth and VerdePilot are built on it. We integrate it into our clients' governed-AI projects, with audit and support. DebugABot remains a research project.
Fill in the form: your request goes to the dedicated team.