# H/M Agent Drift

Canonical page: <https://hogarmas.net/agent-drift/>

H/M Agent Drift is a free, deterministic, browser-local release comparison for declared AI-agent authority. It compares two compatible H/M reports and identifies capability expansion, control removal, control-coverage regression, attack-path changes, and selected Agent Scope aggregate changes.

## Accepted inputs

- `hm.agent-risk-atlas.v1`
- `hm.agent-scope.v1`
- JSON only, up to 500 KB per local file

The baseline and candidate reports are parsed and hashed in the browser. No report is uploaded to H/M. The comparator does not call a model and does not execute a tool.

## Comparison model

1. Normalize each compatible report to capabilities, declared controls, active attack paths, and source-level aggregate counts when available.
2. Compare additions, removals, new paths, resolved paths, and coverage direction.
3. Evaluate only the explicitly selected release-policy gates.
4. Export a source-minimized comparison packet with two local file hashes.

Agent Scope inputs are normalized from their declared tool capabilities. Tool names are used only while parsing locally and are not retained in the normalized comparison. Agent Risk Atlas inputs use the declared capability and control profile contained in the export.

## Decisions

- `review-required`: at least one selected policy gate is violated.
- `change-observed`: normalized changes exist and no selected policy gate is violated.
- `no-observed-drift`: no normalized change is observed under this ruleset.

None of these decisions means safe, approved, certified, or complete. The review-priority points are an ordering heuristic, not a risk score.

## Policy gates

The default policy pack contains:

- `no-new-critical-paths`
- `no-sensitive-authority-expansion`
- `no-control-removal`
- `no-coverage-regression`
- `no-isolation-growth`

The optional strict gate is:

- `same-measurement-method`

Machine-readable definitions: <https://hogarmas.net/agent-drift/policy-packs.json>

## Export contract

Schema: <https://hogarmas.net/agent-drift/schema.json>

The JSON and Markdown exports contain normalized deltas, selected gate statuses, limitations, and the SHA-256 hash of each local input. They do not contain raw reports, source payloads, source filenames, or tool names.

## Privacy and metrics

- Source reports uploaded: false
- Source payloads stored: false
- Source filenames stored: false
- Tool names stored: false
- Model calls: 0
- Tool executions: 0
- Synthetic samples are labeled and counted separately from manual comparisons.

Aggregate product metrics use fixed action identifiers only. Metrics never contain capabilities, controls, paths, differences, source hashes, report contents, filenames, or tool names.

## Explicit limitations

- Runtime behavior is not observed.
- Declared controls are not tested or verified.
- Source integrity and deployment identity are not verified.
- Different schemas or normalization methods can reduce comparability.
- The finite path library cannot guarantee complete threat coverage.
- No security verdict is produced.

## Primary references

- NIST, monitoring deployed AI systems: <https://www.nist.gov/news-events/news/2026/03/new-report-challenges-monitoring-deployed-ai-systems>
- Model Context Protocol security best practices: <https://modelcontextprotocol.io/docs/tutorials/security/security_best_practices>
- NIST, identity and authority for software agents: <https://www.nist.gov/news-events/news/2026/02/new-concept-paper-identity-and-authority-software-agents>
- Model Context Protocol tool annotations: <https://blog.modelcontextprotocol.io/posts/2026-03-16-tool-annotations/>
- CycloneDX machine-learning BOM guidance: <https://cyclonedx.org/guides/OWASP_CycloneDX-Authoritative-Guide-to-AI-ML-BOM-en.pdf>

H/M Agent Drift does not claim CycloneDX compliance.