Use Case · Runtime Protection

Runtime That Catches What Static Analysis Can't Reach.

An eBPF-class agent watches syscalls and library calls in production, scored by anomaly models that already know your code graph. Telemetry is signed and pinned to the deployed artefact — so the post-incident question is "which build", not "which guess".

eBPF
Agent Substrate
<2%
Steady-State CPU Overhead
Signed
Per-Event Telemetry
Code-Graph
Anomaly Grounding

Static Analysis Stops Where Production Begins.

Pre-deploy scanning catches what is visible at the code-and-config layer. It cannot catch the supply-chain library that activates a callback only under specific runtime input, or the misconfigured sidecar that exfiltrates secrets only when a particular feature flag is on.

Legacy runtime tools — EDR / agentless EPP — score behaviour against generic anomaly baselines. The signal-to-noise ratio is poor because the baseline does not know which syscalls are normal for your shipped binary.

The Safeguard agent ties runtime telemetry to the deployed artefact's call graph and SBOM. A syscall that is impossible for the deployed code is a guaranteed alert, not a probabilistic one.

01

Behavioural Activation Hides Until Runtime

Supply-chain attacks frequently include dormant payloads. They are invisible to static analysis and only fire under specific inputs in production.

02

Generic Baselines Are Noisy

EDR tools that learn a baseline from observed traffic alert on every uncommon syscall. Your engineers stop reading after the second false positive on a Tuesday.

03

No Provenance On The Alert

When an alert fires, the team has to map the affected container back to a git SHA and dependency graph manually. The investigation starts behind.

04

Detection Without Code Context Misses Causation

Knowing process X opened socket Y is not enough. Knowing that no path in the deployed binary reaches socket Y is the actionable signal.

What It Does

eBPF Telemetry, Code-Graph-Aware Scoring.

eBPF-Class In-Kernel Probes

Lightweight in-kernel agent collects syscall, network, file, and library-load events with under 2% steady-state CPU overhead on supported kernels.

Per-Artefact Behavioural Baseline

Baseline is derived from the deployed binary&apos;s call graph and SBOM — not learned blindly from traffic. Syscalls impossible for the deployed code never make the baseline.

Eagle Anomaly Scoring

The Eagle ranker scores each anomalous event against the artefact baseline and historical tenant traffic; high-confidence events page on-call.

Signed Telemetry

Every emitted event is signed with the agent&apos;s workload identity and the artefact digest; tamper of in-flight telemetry is detectable downstream.

The Pipeline

From Syscall To Verified Alert.

01
Agent install

Daemon-set deploys across the cluster; node attestation pins the agent to a tenant identity and the deployed artefact digest.

02
Baseline lift

Per-artefact baseline derived from the call graph, SBOM, and the first 24h of observed traffic; impossible-syscall set published to the in-kernel filter.

03
Live event capture

eBPF probes capture syscall, network, file, and library-load events; ring-buffer batched and signed before egress.

04
Eagle scoring

Events scored against the per-artefact baseline + tenant history; high-confidence anomalies open a finding with full provenance attached.

05
Auto-quarantine option

Tenant policy can elect auto-quarantine for confirmed RCE-shaped anomalies; otherwise pages on-call with a one-click isolation action.

06
Post-incident replay

Signed event log re-derives the timeline for the post-mortem; pinned artefact digest links the runtime evidence to the source commit.

Outcomes In Production.

High-Precision Alerts

Per-artefact behavioural baseline
Impossible-syscall alerts are deterministic
Eagle scoring on the borderline cases

Tamper-Evident Telemetry

Workload-identity signed events
Pinned to the deployed artefact digest
Post-incident replay is trustworthy

Tight Containment

Optional auto-quarantine policy
One-click isolation from the alert
Signed quarantine action in the audit log

Tie runtime evidence to Safeguard Code for the call-graph join, incident-response for the campaign loop, and self-healing containers for auto-quarantine targets.

See The Syscalls Your Code Cannot Make.

Deploy the daemon-set in a staging cluster and we'll generate the per-artefact baseline plus the first anomaly report inside a day.