Authorization and blast radius
Bravos drives a coding agent that can send real exploit traffic. This page is what it may do, to what, and on whose say-so.
The hard rules
Section titled “The hard rules”bravos guide safety prints them:
THE HARD RULES
Authorization No request goes to a host other than loopback without a committed, reviewed scope.yaml and your own typed authorization naming the target.
Destructive tests DoS and other destructive classes run only against a disposable loopback target.
Credentials Never invented. Test users and sessions are seeded or captured for real, or the run waits for you.Truncated after credentials; two more rules follow. The agent in your session
writes and reviews the probes; Bravos tests them with cxg scan. It never
spawns another agent CLI or uses cxg model generation. Agents never write
@accepts or @entitles, and never confirm a claim without reproduction
evidence. The effort dial changes how much work a skill does; it changes none
of these rules.
Where a test may go
Section titled “Where a test may go”Loopback is the default. A target anywhere else needs both:
.guardlink/scope.yaml, committed and reviewed.bravos setup --scopewrites a loopback-only starter when the file is absent.- Your own typed authorization naming the target, given to the agent in the session.
Scheduled replay also needs the committed .guardlink/proving.md choice to
be schedule, the exact configured target and reviewed scope, a recorded
human authorization, bound committed kits and the current default-branch
revision. It performs no probe generation and permits no destructive classes.
A missing or revoked grant is not permission to continue.
Automatic loopback replay needs an auto choice on the reviewed default
branch and a recorded target. A one-off run under ask needs the person’s
current authorization. On a clone whose upstream has no testing policy, a
one-time loopback run can use policy and scope committed only in that clone,
with ask and the person’s typed authorization. That grant does not enable
automatic or scheduled replay, and loopback consent covers only the local app.
The helper checks the target build and source/kit
fingerprints again before updating the ledger; it writes receipts while private
HTTP and authentication material stay out of committed evidence.
Proving when autopilot is on
Section titled “Proving when autopilot is on”The first time proving comes up, the agent asks you once how you want it
handled: automatically on your loopback target, ask each time, on a schedule,
or not now. It records your words in .guardlink/proving.md. Say “ask me
first” or “stop auto-proving” at any time to change it. A target other than
loopback still needs a reviewed scope and your typed authorization every time.
Capture real test logins
Section titled “Capture real test logins”bravos-env seeds synthetic users on the authorized loopback app and captures
each required identity separately. For bearer APIs, dynamic login can discover
the token in a real login response and save an Authorization: Bearer profile.
Browser-only login uses a real captured session; probes reuse the profiles
instead of minting tokens. Private credentials and session state stay out of
git. The flags belong to cxg authentication.
What Bravos writes
Section titled “What Bravos writes”In the repository, bravos setup writes the skills, the hooks under
.bravos/hooks/, the hook registrations in each agent’s settings file, and
the MCP entries. All of them are meant to be committed. .bravos/skills.lock
records the sha256 of every file Bravos wrote, so a later bravos setup
updates only what it wrote and keeps a file you edited unless you pass
--force. bravos uninstall removes its unchanged files, hook entries and
.bravos/; --dry-run lists them first.
Results it gathers, such as SAST runs under .bravos/sast/, the dashboard and
report evidence under .bravos/reports/, are gitignored, each folder by its
own .gitignore. Design documents under .bravos/context/ are committed. See
Share Bravos with your team.
On the machine, ~/.bravos (or $BRAVOS_HOME) holds the licence, the staged
analyzer and code graph, the graphs built, and the workspace registry.

