aws-samples / aws-samples/sample-autonomous-cloud-coding-agents
RFC: Runtime security telemetry below the agent process
- Dominant language
- TypeScript
- Stars
- 143
- Forks
- 46
- Avg merge
- 3d 9h
- Merged PRs (30d)
- 20
Description
## Primary area
** CDK / infrastructure
## Related issue or feature request
_Complements roadmap **Behavioral circuit breaker**, **Workload-anchored credential binding**, and **Immutable audit export**._
## Summary
Explore integrating **host-level runtime telemetry** (syscall/file/network events from outside the agent process) into ABCA's audit trail, so operators can corroborate or contradict agent-reported actions during incidents. Cedar and application hooks remain; this adds a physics-layer signal.
## Use case and motivation
Application-level guardrails can be reasoned around when instructions and tools coexist in the same trust domain. Editable JSON allowlists are trivially tampered with; **kernel-visible truth** (syscall/file/network) corroborates agent-reported progress and supports adaptive containment (e.g. credential exfiltration patterns). Security teams want tamper-evident records of what the sandbox actually did—especially for unattended background tasks—without relying solely on LLM narrated progress events.
## Proposal
### Architecture sketch
1. **Collector** in MicroVM guest or sidecar streams normalized events (file open, exec, connect) to CloudWatch Logs / S3 with `task_id` dimension.
2. **Correlator** Lambda joins runtime events with `TaskEvents` on `task_id` + time window for operator UI / `bgagent trace`.
3. **Policy feedback (optional later):** high-risk syscall patterns trip **Behavioral circuit breaker** even if Cedar allowed the tool call.
### Deployment stance
- Phase 0: design doc + spike on AgentCore Runtime observability hooks available today.
- Phase 1: opt-in Blueprint `runtime_telemetry: true` for pilots.
- No default-on until performance and PII review complete.
## Out of scope
- Shipping a specific third-party agent product integration in core repo.
- Replacing Cedar decisions with kernel rules in v1.
- Windows/macOS dev environments (compute target is Linux MicroVM only).
## Potential challenges
- AgentCore abstraction may limit BPF/eBPF access; may require alternative compute strategy for full fidelity.
- Volume and cost of raw syscall logs.
- PII redaction in paths and env vars.
## Dependencies and integrations
- `COMPUTE.md`, `SECURITY.md`, `OBSERVABILITY.md`
- Operator dashboard / `bgagent trace download`
## Alternative solutions
- **Stronger Cedar only** — necessary but insufficient when bash and subprocess paths can bypass declared tool calls.
- **Post-task filesystem diff only** — lighter weight; misses network exfiltration.
---
**Note:** Non-triaged RFCs may not get timely review. PRs on non-triaged issues might not be accepted.
* RFC PR:
* Approved by:
* Reviewed by:
Contributor guide
Research direction
Start by reading COMPUTE.md, SECURITY.md, and OBSERVABILITY.md, then inspect the AgentCore Runtime observability hooks referenced in the proposal. Determine whether a design document and feasibility spike can capture host-level telemetry, correlation, deployment, cost, and PII constraints. Done means a reviewed Phase 0 design and spike, not a production integration.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- aws
- Domain
- cloud, infrastructure, observability, security
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Quiet
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100