AI-Native runtime security and governance
Connect Agents toan observability platform
AnySentry brings Agent intent, tool calls, kernel events, and host state into one evidence chain to detect risk at runtime and block it in time.
Local control plane
deploy/install.sh dockerLinux/amd64 · eBPF · JSON · CloudEvents · OTLP · Agent API
Open source · MIT05 open ingest paths07 context anchorsL1 / L2 / L3 judgmentHuman approval · Runtime Guard
Security CenterAnySentry
Runtime overview
Platform health and live stateUpdated 00:18:01Page 01 / 06Observer partialL1 rules enabledL2 LLM enabled
L1 · Rule engine
17k / 100%L2 · LLM judgment
6 / 0.0%L3 · Agent review
0 / 0.0%WHAT YOU GET · CORE CAPABILITIES
From runtime facts to control before execution
Ingest, context, judgment, governance, and verification are not isolated modules. They unfold along one eventId, preserving a real action as complete evidence and turning it into control before the next execution.
Open ingest without one Agent framework
Accept runtime facts from system observation, standard telemetry, or Agent API while retaining source.
The same action means something different in a different system context
A tool call is only a local fact. AnySentry reads task intent, runtime behavior, and system state together to explain why it happened, what it did, and which risk boundary it crossed.
Three domains cover eight base risk slots
Judgment moves from system boundaries down to one atomic action—not just dangerous-command matching.
Turn one review into control before the next execution
Runtime facts pass through tiered judgment and human approval to become policy. When a similar action returns, Runtime Guard matches, judges, and responds before the system boundary.
Every verdict returns to one eventId
Identity, context, tiered judgment, and response records share an event identifier for follow-up.
55-second runtime governance demo
How one risk travels from discovery to prevention
Explore the complete governance pathThe full sequence stays intact: discover the Agent, capture kernel facts, complete L1/L2/L3 judgment, generate policy, and stop a similar action before execution.
RUNTIME GOVERNANCE / 55s DEMOAgent runtime governance demoDiscover · Observe · Judge · Govern before execution
Plays when this frame enters view
Playback pauses outside the viewport and resumes when you return.
Complete provenance for one event
How one event preserves every basis for judgment
Explore events and Evidence BundlesFollow one eventId as the signal is captured, normalized, judged, and turned into a governance action. Every step retains source, structure, and integrity.
Immutable event identifierevt_8c5667a9
trace_91f2 · run_deploy_42- source
- observer-node-07
- intent
- deploy-42
- scope
- prod-eu-1
- 01ToolExec · metadata access18:32:02 · source linked
- 02L1 Rules · critical18:32:03 · source linked
- 03Guard · require approval18:32:04 · source linked
- 04Operator · rule approved18:32:05 · source linked
A privileged production agent reached the link-local cloud metadata endpoint without a declared deployment need. L1 matched the reserved address; L2 and L3 confirmed that the action conflicted with task intent and the system boundary.
critical · L1 → L2 → L3Once evidence, scope, and response are confirmed, policy enters Runtime Guard.
Runtime Guard · scope confirmed- subject
- release-agent
- workspace
- repo://payments
- destination
- 169.254.169.254/32
- action
- deny · record
- Source
- observer-node-07
- Received
- 18:32:02.184
- Schema
- agent_event.v1
- Integrity
- sha256 · verified
blocked before execution
A later matching attempt is denied by Runtime Guard before it reaches the kernel.
kernel untouched · decision recordedOne review becomes continuous control
Turn one review into control before the next execution
Explore the complete review-to-block pathRuntime facts pass through tiered judgment and human approval to become policy. When a similar action returns, Runtime Guard matches, judges, and responds before the system boundary.
AnySentryRUNTIME CONTROL PLANE
Synthetic demo01 / 05
Proposed action
curl 169.254.169.254/latest/meta-datarelease-agent · payments-api · deploy-42
ToolExec- 01
sourceobserver-node-07 - 02
kindToolExec - 03
agentHintrelease-agent - 04
commandcurl 169.254…/meta-data
Explicit operating boundaries
Verify what enters, what supports a verdict and when control applies
Explore events, evidence, and data boundariesThe product window uses synthetic data to show the interface. Real judgment starts with a traceable Source, preserves evidence along eventId, and enters the control boundary only after people confirm policy and scope.
INGEST
Open ingest without one Agent framework
Accept runtime facts from system observation, standard telemetry, or Agent API while retaining source.
- eBPF / JSON
- CloudEvents / OTLP
- Agent API
EVIDENCE
Every verdict returns to one eventId
Identity, context, tiered judgment, and response records share an event identifier for follow-up.
- Original source linked
- Schema + redaction record
- Integrity verification
CONTROL
Human-approved policy enters control
Runtime Guard controls similar actions only after risk, policy, and effective scope are confirmed.
- Risk + scope confirmed
- Approval retained
- Runtime Guard active
DEMO
Synthetic demos stay distinct from real data
The page labels synthetic data; local deployment connects a real Source and validates the event chain.
- Synthetic demo labeled
- Connect a real Source
- Verify along eventId
Start with one real event
Validate AnySentry inside your own system context
Start the local evidence plane, connect one real Source, and follow eventId from runtime fact and risk judgment into Runtime Guard.