docs/security/.
It was done before submission with Claude Code (Claude Opus 5.5 for the review, Claude Fable 5.2 for the fixes, the threat model and the attack test), and a human approved every change. It is an internal audit: no outside party has audited HelloNear. It has three parts:
- Review of the repository and the hosted API: leaked secrets, what the public API gives away, and ways to exhaust the shared NEAR AI credit or overload the API (the two tables below).
- Threat model of the hosted demo: 16 threats mapped to STRIDE and the OWASP Top 10:2025, as a Threagile model that runs (
threagile.yaml), with Sigma detection rules for the high ones. - Attack test from outside against
api.hellonear.xyz: about 220 requests and 2 runs (the third table below).
Each fix has tests (
limits for every user in apps/be/test/api.test.ts, engine/test/autopilot-limits.test.ts) and ran under the load tests before merge. The limits belong to the hackathon phase.
The attack test found one real gap (the IPv6 row above), which is fixed. It also showed what is still missing: these refusals are not written to a log the team can query, because the backend logs only its
5xx answers. An access log is planned for after judging.
The hosting side was checked by the team (#327): the box’s firewall lets only the tunnel host reach the app ports, so the per-IP limits see the real visitor; Cloudflare’s DDoS protection and managed ruleset are on; and an edge rate limit blocks an IP for 10 seconds after 60 API requests in 10 seconds.
Known limits: scheduled Autopilot runs do not count toward the daily total for all users; they are capped at 100 a day per owner by the engine. A client with both IPv4 and IPv6 gets two per-IP allowances. Requests the API refuses (4xx) are not logged yet. The smart contracts were not reviewed beyond their existing tests. No outside party has audited HelloNear.
Check it yourself
Two refusals anyone can see without an account. The first prints401 (no session), the second 403 (a request from another origin):