Skip to content

Security: BrightbeamAI/brevet

Security

SECURITY.md

Security Policy

Reporting a vulnerability

Please report suspected vulnerabilities privately to oss@brightbeam.com. Do not open a public issue for security reports. You will receive an acknowledgement within five working days.

Reports of particular interest, given what Brevet protects:

  • any way to raise a capability's authority without a recorded human or mission-group identity (invariant I2 in SPEC.md section 2)
  • Evidence-layer material resolving into a lockfile (invariant I1), or recalled material reaching a lockfile or a running agent (SPEC.md section 7)
  • a decision, release or register change accepted without the approver signatures the workspace requires, including a change to a mission group made without that group's quorum
  • evidence-chain manipulation that brevet verify fails to detect, including a chain cut short, rewritten or replaced behind its anchors
  • an agent running, outside the shadow channel, anything other than its latest signed release: a modified manifest, lock or harness that still verifies
  • a tool call that escapes the tool tiers a release declares

Supported versions

Version Supported
0.4.x yes
0.3.x and earlier no: upgrade to 0.4

Scope notes

The rules a conforming deployment enforces are in SPEC.md; how each control is switched on is in ABOUT.md under "Signed approvals" and "Running in production". Approver keys live outside the workspace, encrypted with a passphrase or held as the approver's own SSH key, and can be checked against an allowed-signers file or GitHub. Where anchors are configured, the evidence chain's head is anchored outside the workspace, in a file the agent cannot write or a CHAP coordinator's audit log. Brevet assumes an attacker who controls the workspace does not also hold an approver's unlocked key or write access to every anchor target.

There aren't any published security advisories