bravos

Shift-left security that runs with every commit.

Bravos makes the coding agent you already use work your threat model, code graph, SAST and authorized pentests, then gates every pull request and keeps a test that guards each fix.

$ npm install -g @bugb-technologies/bravos

Open your project in your coding agent (Claude Code, Codex, Cursor…), then paste this.

Full install guide Talk to sales

Add order export pinecart · #482running

+1 exposure declared on the new route: #orders-api → #idorEvery risk mapped, with 5 still open

02 · Threat modelThreat modelling

Every risk, mapped where it lives.

Written next to the code, drawn four ways.

Assets 8Open 5Mitigated 7Controls 7

12 of 12 exposures and 7 mitigations drawnopen, by severitymitigatedhover or tab to tracepinecart · sample

5 open7 mitigated8 assets

Threat graphEvery risk tied to the part of the app it threatens, and the fix that covers it.

21 flows14 cross a trust boundary

Data flowData traced from the outside world, through your app and out again.

4 zones14 crossings, each marked

Trust zonesWhere data moves from one zone of trust to another, the crossing is marked.

4 direct8 two hops out

NeighbourhoodPick one asset and see everything it talks to, two hops out.

1 / 4

03 · Code graphTeam plan

Your codebase as a map. Every route to what it touches.

Lens

Where in this codebase is the risk?

5Open exposures5 files

7Controlsdeclared mitigations

21On a flowdeclared data flow

4Hotspotsby commit heat

18Entry pointsroutes and tools

8 layers · 20 folders · 377 filesopen exposure, by worst severityhover or tab to a folderpinecart · sample

Api

Agent

Workers

Data

Web

Lib

Test

Config

377 files5 folders with an open risk

MapYour whole codebase on one screen. Risk shows up as heat.

15 pairs11 with no import between them

CouplingFiles that change together, linked, even where no import joins them.

18 ways in11 reach a risky call

Attack surfaceEvery way in, ranked by the risky calls it can reach.

2 unapproved (dashed)1 needs a person's sign-off

Agent reachWhat your AI assistant can do, and which of it nobody approved.

1 / 4

04 · SASTTeam plan

Findings, sorted by what can reach them.

Pre-PR gatefail

Fail the gate0

Warn0

On a path from an entry0of 61

Deterministic analysis, placed on the code graph. Nothing is sent anywhere.

Issues by evidence tier

T1 proven0, 0 reachable

T2 corroborated7, 4 reachable

T3 plausible11, 5 reachable

T4 hygiene29, 10 reachable

T5 set aside14, 4 reachable

Narrowed by reach

61all findings
23on a path from an entry point
9corroborated or plausible
2fail the gate

Issues

User-controlled URL reaches http.getsrc/uploads/import.ts:31via POST /uploads/importfail@exposes #uploads to #ssrf
No ownership check before db.orders.findAllsrc/api/orders.ts:42via GET /orders/exportfail@exposes #orders-api to #idor
Unescaped HTML rendered in admin notessrc/admin/notes.tsx:66via POST /admin/noteswarn@exposes #admin to #stored-xss
Model output passed to a tool uncheckedsrc/agent/tools.ts:19via tool: send_emailwarn@exposes #support-agent to #prompt-injection
Signature compared with ==src/webhooks/payments.ts:14via POST /webhooks/paywarn
Path built from the upload filenamesrc/uploads/store.ts:27via POST /uploads@mitigates … using #path-normalise
Secret read from env written to a logsrc/admin/settings.ts:18no entry reaches itwarn@exposes #admin to #secret-leak
Dependency with a published advisorypackage-lock.jsonno entry reaches it
Hard-coded key in a test fixturetests/fixtures/pay.ts:3no entry reaches it
Regex with a nested quantifiersrc/lib/slug.ts:9no entry reaches itreview: noise

showing 10 of 61 · sample

T1 provenT2 corroboratedT3 plausibleT4 hygieneT5 set asidepinecart · sample

Narrowed by reachSixty-one findings, narrowed to the two that need you today.

23 of 61 on a path from a way in

Placed on the code graphEach finding comes with the route that reaches it, and the risk it matches.

Deterministic. Nothing is sent anywhere.

The gateThe two that matter fail the pre-PR gate and hold the pull request. The rest warn or wait.

1 / 3

05 · ProvenAuthorized testing

Then it tests a running copy, and keeps the proof.

Against a target you own, with real logins, inside the scope you committed.

This run, accounted for

12 tracked · 0 reached a verdict

0confirmed0ambiguous0not run0refuted

The proof it keeps

0request and response pairs, one per confirmed finding

0retest kits kept in your repository, ready to replay

0@confirmed lines, recorded in the project's evidence ledger

0probes, every one inside the scope you committed

“Not run” and “ambiguous” are listed, not hidden. They prove nothing either way, so they stay untested.

3 logins · one run · 47 probesnumbers only · sample

47 probesevery one inside the scope you committed

Three logins, one runIt tests a running copy of your app that you own, signed in as each kind of user.

2 confirmed1 ambiguous1 not run8 refuted

Every risk accounted forEach tracked risk lands in a bucket. What didn't run, or ran without a clear answer, is listed, not hidden.

2 proofs2 retest kits saved

The proof it keepsEach proof is kept: the request, the response, and a retest kit that replays it.

1 / 3

06 · Gate and retest

Each proof becomes a test that guards the fix.

A retest kit in your repository, and a gate on every pull request.

The two retest kits stay in your repository. /bravos-recheck replays them after a fix; the nightly can replay them against a target you authorize.

Fix order ownership and import allowlist#487

src/api/orders.ts

41  const orders = await db.orders.findAll({42+   where: { customerId: session.user.id },43  })

Bravos #orders-api → #idor — confirmed, on a changed path

bravos security gate: failed

Passed: Threat model parses and resolvesModel validated
Passed: No undisclosed new exposurenothing new
Failed: No confirmed exploit on a changed path2 on changed files

Retest kits, replayed by /bravos-recheck

severity	asset → threat	verdict	location	outcome
2 confirmed · 0 refuted of 12 tracked
high	#orders-api → #idor	confirmed · kit	src/api/orders.ts:42	still exploitable
high	#uploads → #ssrf	confirmed · kit	src/uploads/import.ts:31	still exploitable

▸ bravos security gate  ✓ Threat model parses and resolves  ✓ New claims meet the evidence bar  ✓ No undisclosed new exposure  ✗ No confirmed exploit on a changed path    Confirmed-exploitable claims on changed files, with no @ticket for them in that file:    high #idor on #orders-api — src/api/orders.ts:42    high #ssrf on #uploads — src/uploads/import.ts:31

Where it shows up

Your session/bravos-ship makes the same rulings before gh pr create

Beside your agentbravos watch: open, confirmed, tested and stale, live

GitHubthe gate's check on the pull request, and SARIF in code scanning

Before the fixBoth proofs still work. The gate fails a pull request that leaves a confirmed exploit on a changed file.

After the fix/bravos-recheck replays both retest kits. Both attacks are blocked, the ledger records them refuted, and the gate passes.

2 retest kits, in your repository

Every weekday nightPointed at a target you authorize, the nightly replays both kits, so the bug can't quietly come back.

1 / 3

How it fits

What Bravos provides, and what you bring.

Constant · provided by Bravos

Shared context

The state every skill reads and writes back. It lives in your repository, shared through git.

Threat model
Hypothesis ledger
Retest kits
Code graph

skills /bravos-…contextannotatecodegraphsastreport

Constant, provided by Bravos
Changes over time, you bring it

01AnnotateThe threat model, written into your code./bravos-annotate
02TemplatizeEach open exposure becomes a testable security check./bravos-templatize
03ProveProbes run against a target you own, with real logins./bravos-prove
04LedgerEach result goes into the project's evidence ledger; kits replay after a fix./bravos-recheck
05GateA confirmed exploit on a changed path fails the pull request.bravos sync --ci

bravos

Secure your next pull request.

Install Bravos Talk to sales

$ npm install -g @bugb-technologies/bravos

Install as a plugin in your coding agent

See the plugin guide for supported coding agents, installation instructions and differences from CLI setup. Every plugin needs the CLI above.

Only want the pull-request gate? uses: Bugb-Technologies/bravos-action@v1 needs no licence. GitHub Marketplace

See licences