Built around the endpoint. Verified at every hop.
Vedric is an agent, a cloud, and an operator surface - connected by signed messages and tenant-isolated at every layer. Nothing collected is content. Nothing executed goes unlogged. Nothing written can be quietly undone.
Three layers. One signed channel between them.
The agent
Runs as a managed service under OS-sealed credentials. Collects behavioural signals from the host, identity, and network layers; the specific collector mix is not documented publicly. Content is never read, never stored, never shipped.
The correlator
Baselines behavior per user and per peer group, correlates deviations into storylines, and runs triage against a strict budget. Every query is tenant-scoped. If the AI layer is unavailable, baseline detection keeps running.
The surface
Your analyst sees one alert per correlated storyline - not a flood of events. Every response action is admin-gated, requires a reason, flows as a signed command to the endpoint, and is recorded to an audit trail that cannot be quietly edited.
The agent
Runs as a managed service under OS-sealed credentials. Collects behavioural signals from the host, identity, and network layers; the specific collector mix is not documented publicly. Content is never read, never stored, never shipped.
The correlator
Baselines behavior per user and per peer group, correlates deviations into storylines, and runs triage against a strict budget. Every query is tenant-scoped. If the AI layer is unavailable, baseline detection keeps running.
The surface
Your analyst sees one alert per correlated storyline - not a flood of events. Every response action is admin-gated, requires a reason, flows as a signed command to the endpoint, and is recorded to an audit trail that cannot be quietly edited.
How a deviation becomes a signed verdict.
- 01 /CollectEndpoint behavior is collected across several surfaces by default. Deeper visibility modes can be enabled per tenant. What is never collected is documented and refused in code.
- 02 /NormalizeUser identity is anonymized on the endpoint. The backend stores a one-way token - not the raw identity. Your tenant can map it back; we cannot.
- 03 /DetectTwo layers run in parallel. Behavioral baselines (per user, per peer cohort, per endpoint) surface drift from the user's own pattern. Rule-based detection catches well-known-bad patterns that fire from day one. A quiet warm-up prevents the baseline layer from generating noise while it settles; the rule layer is active immediately.
- 04 /CorrelateDeviations and rule hits on the same endpoint and user within a short window become a single correlated storyline.
- 05 /TriageTriage runs inside a fail-closed budget and circuit. It reads tenant-scoped data only and never reaches the endpoint. If it is unavailable for any reason, baseline detection continues.
- 06 /RespondResponse actions - isolate host, kill process, quarantine file, disable session - are admin-gated, require a written reason, and flow as signed commands. No anonymous or unreviewed actions exist.
- 07 /ArchiveEvery decision lands in an append-only audit. Edits and deletes are refused at the storage layer. Your forensic trail is preserved without depending on operator discipline.
Six constraints the product is built around.
Metadata-only collection
The refusal list is enforced in the agent source code, not in policy. Content cannot be quietly added by changing a config; the product's collection surface is constrained by design.
Tenant isolation by default
Every tenant-scoped query is automatically filtered. The guarantee is tested on every build - if a new table would create a cross-tenant path, the build fails.
Offline-signed updates
Agent binaries are signed with a key that never lives on the server. A cloud compromise alone cannot produce a valid update; the attacker would need offline key custody too.
Append-only audit
Every audit surface refuses edits and deletes at the storage layer. An operator with direct data access cannot quietly remove an entry - the write itself is refused.
Fail-closed AI
Triage runs inside a strict per-tenant budget and a fail-closed control layer. An operator kill-switch propagates across the platform quickly. Baseline behavioral detection continues regardless - the AI layer is useful, not load-bearing.
Signed response actions
Every response action requires admin role, a written reason, and flows as a signed command. There is no path to an anonymous or unreviewed action anywhere in the product.