# Security scope: what Bucker will not do about your security

**Draft, awaiting the owner's ratification of version 1.0.** The decision behind it (OD-DP-19) is
open, and the ratification is recorded against a version of this page; none is recorded yet.

Bucker is an application observability and remediation platform. Its security work is about the
application it already watches: attacks visible in the requests that reach your code, findings in
your dependencies and code, and the audit trail of what people and agents did inside Bucker. This
page is published before most of that ships, so the boundary is on record and nobody has to infer it
from silence. It is version 1.0, as of 2026-11-04.

## What Bucker does in this area

- **Reads attack signals from application telemetry.** Requests that carry exploit patterns are
  classified by a published rule set, and an issue that matches enters a *security lane*: no
  automatic pull request, a sealed sandbox, and a human review before the lane lifts.
- **Surfaces findings as ordinary issues**, each with the threshold that produced it, so a finding
  can be triaged, assigned and resolved like an error.
- **Proposes, never applies, blocking.** A block rule starts in monitor mode and is staged by a person.
- **Forwards its own audit and security events** to your SIEM instead of becoming one.

## Acts only a person performs

These cannot be requested by an agent, a service principal or a custom role, on any door. The
registry marks each one human-only, and the engine, the MCP gate and the REST routes each refuse a
non-human before they look at policy, org settings or the agent's state.

| Act | Who | Why a person |
| --- | --- | --- |
| Lift or confirm a security lane (`threat.lane.review`) | A person holding `approve_fix` on the issue | An agent cannot clear the very flag that restrains agents. A confirmed lane is permanent. |
| Change the organization's security detection policy (`security_detection_policy.update`) | An owner or admin | What counts as an attack is the organization's decision. |

## What is not offered

Each item has no action name, so no policy row can enable it. *Kept out by* names the line or
decision that keeps it out.

| Not offered | Why | Kept out by | Instead |
| --- | --- | --- | --- |
| A SIEM: log detection engines, rule modes, SQL detections, content packs, threat-intelligence feeds, an IOC explorer | A SIEM is a different product with its own meters and operators. | OD-DP-19 | Bucker forwards its own audit and security events to yours. |
| Cloud security posture (CSPM), Kubernetes posture (KSPM), cloud identity risk (CIEM), attack-path graphs, cloud resource catalogues | Each needs standing access to your cloud accounts or clusters. | DNB 2, OD-DP-19 | Findings about your code and dependencies, mapped to the code that owns them. |
| Remediating running cloud resources, including AI-suggested fixes to misconfigurations | Bucker does not hold credentials that change your infrastructure. | DNB 2 | Code fixes ride the existing remediation lane, where a person approves the pull request. |
| Runtime and workload protection: kernel or eBPF agents, file integrity monitoring, automated process kill or network isolation | They need an agent on your hosts and the power to act on them. | DNB 4 | Application-level signals only. |
| User and entity behaviour analytics over your cloud and identity-provider users | A risk score over identities Bucker does not hold. | OD-DP-19 | Bucker baselines its own agents and OAuth clients. |
| Per-event, per-GB or per-host metering of security analysis | Findings are not billed by volume analysed. | OD-DP-19, OD-DP-06 | Security findings are unmetered. |
| Automated blocking that has not been staged by a person | A rule that blocks real traffic changes production. | F-11 | Block rules start in monitor mode and are promoted by a person. |

## Where this is enforced

The human-only acts above are rows in the action registry. A registry-wide test walks every
human-only act and proves the engine denies it to an agent and to a service principal, whatever
policy, organization flag or agent state surrounds it, and that no custom role can be granted the
scopes behind them. `reviewIssueThreat`, the function that lifts a lane, asserts the registry row
before it does anything else, so editing the row away fails closed instead of silently allowing the
act. See [Agent governance](agent-governance.md) for how requests are decided and
[Cost governance](cost-governance.md) for the equivalent boundary around spend.
