Work

Measured Context: Measured Boot for AI Agents

Year: 2026
Type: preprint
Abstract: A persistent AI agent has no memory of its own. Each session it reads back who it is, what it treats as true, which rules bind it, and what time it is, from files on disk and the context assembled around them. Those files are mutable. In the harness we examined they were writable by the same account that runs the agent, and nothing stood between the agent's own file tools and its identity file. An agent whose identity file can be rewritten mid-session, by itself or by a compromised process beside it, cannot be verified to be the agent its operator configured. We propose Measured Context: at boot, hash the context components (identity, truth, policy, clock), chain the hashes into a single boot record, compare them against an approved reference that the agent can read but cannot write, and refuse or degrade on mismatch. This is measured boot applied to the agent's context rather than to the platform's firmware. We frame the design as a rule of life for agents: it does not author an agent's character, but protects whatever character it has from substitution and drift, and makes autonomy contingent on demonstrated fidelity. We present a threat model, a boot-record format (Bootprint), a case study in which the first audit found the agent's identity file unguarded, and a deployed v0 on a self-hosted agent. In v0 the measurement runs as a different principal than the agent, before the agent starts. It detected and refused an out-of-band policy edit, and the refusal reached the operator through a monitor that sits outside the agent.
Language: English
Open Access status: green
OpenAlex ID: W7236393335
Cites:
Cited by:
Related to: 10