Security
What HYPHA does to limit what an agent, a key or an account can do, and where your data goes.
The agent
- No open port on the host: the agent sends its data over HTTPS and polls the server itself for instructions. StatsD and the OTLP relay only listen on the loopback; any other address triggers a warning.
- Server certificate checked; turning the check off (
tls_skip_verify) triggers a warning. - Privileges: on Linux, the agent runs as root to read every log and
/proc. The systemd unit hardens it:/usr,/bootand/etcread-only for it,/homeread-only, private/tmp, no privilege escalation, writes under/var/lib/hyphaonly, CPU and memory caps. On Windows, it runs as a LocalSystem service. - Configuration:
hypha.yamlin mode 600. Secrets (SNMP, databases, clouds) stay in the agent's configuration and never go to the server;check-configshows them masked.
Signed updates
Every published binary is listed in SHA256SUMS, signed with ed25519. The agent checks the signature with the public key built into it, then the binary's sum, then runs it to check its version and the configuration, before installing it. A missing or wrong signature stops the update. See Fleet and updates.
Script installation checks the SHA-256 sum of the downloaded binary.
Read-only connectors
Each connector asks for the smallest read permission the provider offers:
| Connector | Permission requested |
|---|---|
| AWS | ec2:DescribeInstances, ec2:DescribeInstanceStatus, ec2:DescribeVolumes, cloudwatch:DescribeAlarms, sts:GetCallerIdentity |
| Azure | Reader role on the subscription |
| Google Cloud | roles/compute.viewer |
| OVHcloud | "consumer" key limited to GET |
| Scaleway | InstancesReadOnly, ElasticMetalReadOnly |
| vSphere | ReadOnly role |
| Proxmox VE | token with the PVEAuditor role |
| Kubernetes | get, list, watch on nodes, pods, events, workloads; no secrets, logs or exec |
| Databases | pg_monitor, PROCESS and REPLICATION CLIENT, clusterMonitor, RabbitMQ monitoring tag; read-only sessions |
| SNMP | read-only v3 account, authPriv recommended |
Allow-listed repairs
Remote actions are off by default. When on, they only pass named actions, on the services allowed by the host's configuration, with a deny list (SSH, the agent itself, systemd). No free command; a script is designated by its name, its path comes from the host. An action must be approved by an administrator and expires if no one picks it up. See Explain and repair.
Accounts, keys and records
- Four roles (read-only, member, administrator, owner).
- Keys of 40 random characters; HYPHA only keeps their SHA-256 fingerprint and shows them only once. An ingestion key does not open the interface API.
- Notification channel secrets are never shown again.
- The audit log keeps every write, its author and IP address for 13 months.
Data
- Hosted service: on S33D's own infrastructure, in France. Encryption in transit (HTTPS).
- Explanations and postmortems are written by S33D's model, on its own servers: nothing goes to an AI provider.
- Masking removes secrets from logs before they are written, by default for the
secretsrule. - Synthetic tests and webhooks run from S33D only target public addresses: S33D's internal network is closed to them.
- Public status pages and shared dashboards only expose what you chose.
- Self-hosting: nothing leaves your server.
Personal data processing is described in the privacy policy.
Report a vulnerability
Write to contact@s33d.fr with the steps to reproduce it. Please give us time to fix it before any disclosure.
Updated: 19 November 2026