This repository contains a working participatory-surveillance stack for One Health observations in Arizona. Bugs that affect the confidentiality of observations, the integrity of cluster-detection signals, or the availability of agency-side dashboards during an active outbreak are treated as security incidents — not just functional bugs.
Placeholder private channel — to be replaced at community handoff. The maintaining team will publish a real private disclosure address (a
security@mailbox and/or a GitHub Security Advisory private-disclosure link) as part of the v1.0 handoff. Until then, please use the temporary channel below.
Temporary disclosure address (handoff placeholder):
epihack-security@arizona.edu (placeholder — confirm with the
maintainer before sending sensitive material).
When reporting, please include:
Do not open a public GitHub issue for security reports. Please allow up to 10 business days for an initial response and 90 days for a coordinated fix before public disclosure. The standing review board (see GOVERNANCE.md) is looped in on any report that touches tribal data; ITCA-TEC is notified directly on those.
In scope:
mcp/, agents/,
schema/, app/, dashboard/,
today/, and the deploy config under ansible/./v1/reports endpoint,
the DuckLake (kg_writer) and MongoDB (mongo_writer) sinks, and the
mongo_to_ducklake sync (see plan/09).Out of scope:
These are the five threat classes the repo is designed against, with the mitigations already in place and the gaps that remain.
Threat. An analyst or external party with access to a dashboard or aggregated feed reverse-engineers individual line data — most acutely for observations on tribal land or for low-incidence pathogens (hantavirus, plague) where a single cell can identify a household.
Mitigations in place.
agents/src/onehealth_agents/validation.py) at write
time, before any row reaches the kg.consent_profile defaults to suppress for every new tribe.*
node until an MOU attaches an opt-in (see
GOVERNANCE.md).k and theta floors; see
plan/CLUSTER-CALIBRATION.md.Threat. An API key, agency token, or OAuth refresh credential
for VectorSurv / Twilio / a private agency endpoint is leaked via a
log, an agent_run row, a stack trace, a public bug report, or a
committed .env.
Mitigations in place.
os.environ. No tokens in
source. Each server’s .env.example lists the required vars..env files are .gitignored (see .gitignore at the repo root).agent_run audit rows store input/output SHA-256 digests, not
raw payloads or headers.MONGODB_URI from the environment
(vault-managed at deploy time); when it is unset the writer falls back
to an in-memory mongomock client, so dev and tests need no real
credential. Self-hosted MongoDB is bound to 127.0.0.1 with
authorization enabled and an app user created from the vault password.Threat. A legitimate agency-side credential is reused outside its intended scope — for example, an ADHS token used to write into a tribal-jurisdiction observation, or a CHW field-app token used to read data outside that CHW’s caseload.
Mitigations in place.
consent_profile on every write regardless of caller
identity — and it runs before the per-channel sink (DuckLake or
MongoDB) is selected, so neither channel can bypass it.knowledge-graph-mcp SQL escape-hatch (kg_sql) is SELECT-only:
the server parses the SQL and rejects any non-SELECT keyword
before execution. Do not weaken this filter.vectorsurv_* only reaches
vectorsurv-mcp, etc.), so a misconfigured route cannot leak across
servers.Threat. A user-controlled field on an observation (free-text symptom note, species hint, address) contains LLM prompt-injection content that, when summarised on an agency dashboard or by the Triage Agent, causes the LLM to disclose data, change its triage class, or call a different MCP tool.
Mitigations in place.
plan/03-agentic-architecture.md);
free-form triage output is rejected by the orchestrator before
reaching downstream agents. This is the single most important
injection mitigation in the stack.HEAT_SCORE_TABLE (also in plan/03),
not free-form LLM scoring.dashboard/, today/) render only
pre-aggregated fields from the kg, not raw observation text.value_text and shown only with
HTML-escaping in vanilla DOM updates (no innerHTML shortcuts).Threat. A miscalibrated detector fires on noise, triggering an emergency response (cooling-center surge dispatch, mosquito-control intervention, public alert) that disrupts services without medical justification — and erodes agency trust in the system.
Mitigations in place.
k) and posterior thresholds.
See plan/CLUSTER-CALIBRATION.md.ClusterAlert carries audit fields (tier1_score,
tier2_posterior, rule_tripped, historical_match) so an
analyst can decompose every fire.We document gaps because hiding them does not make them go away. If you can close any of these, please open a PR.
coconino_plague_2025 historical case is the canonical example.
Mitigation gap: no automated escalation today for Y. pestis,
Sin Nombre, or similar pathogens at n=1. Planned for Phase 4.plan/CLUSTER-CALIBRATION.md come from a single fixed RNG seed.
A statistically robust calibration should sweep ~100 seeds.security@ mailbox and a GitHub
Security Advisory configuration.index.html and the per-section
pages load MapLibre and Cytoscape from unpkg without a Content
Security Policy. A future PR should add a CSP header (via Jekyll
_includes) restricting script sources to unpkg + self.NOTICE but do not yet ship a CycloneDX SBOM. Planned.agent_run digest collisions. SHA-256 of canonicalized JSON
prevents accidental disclosure but does not prevent a
determined attacker who has the raw observation from confirming a
match. This is the intended threat model (audit, not
encryption-at-rest) but worth being explicit about.wearable-mcp
server is mock-by-default. The production privacy model
(on-device-only with explicit per-metric consent) is in the README
but not yet enforced at the wire level._photo_has_exif_gps defence-in-depth check
(agents/.../api/routes/reports.py) currently returns False; the
client-side strip (app/src/lib/exif-stripper.ts) is the only
active control today. The server-side check and the full
Triage/Enrichment agent chain land in a later phase
(plan/07, plan/09); until then the contract’s photo_exif_gps_present
(422) rejection is specified but not yet enforced at the wire.mongo_to_ducklake tick. Cluster signals from mobile-only reports
therefore lag by the timer interval. The job is idempotent on
observation_id, so re-runs are safe; a change-stream consumer is the
lower-latency upgrade path.app/src/lib/offline-queue.ts) parks the
already-coarsened, EXIF-free report payload in the device’s
localStorage until it flushes on reconnect, and queued payloads
survive app restarts. That payload includes any free-text notes in
plaintext (the server digests them only on receipt); photos are not
queued. On a shared or compromised device the cache is readable until
the report lands.MONGODB_URI (vault-managed); the in-memory mongomock fallback
persists nothing across restarts and must never back a live deployment.CHANGELOG.md entry).We will credit reporters in the advisory and the changelog unless asked otherwise.