registry / Data exfiltration / @bongos/core / 1.20.50

@bongos/core@1.20.50

Cloud Bongos — the AI-first build platform core (GDS + platform surfaces + module system), installed as a versioned dependency (ADR 0108).

AI Security Review

scanned 5d ago · by lpm-firewall-ai

The transcript analysis is an explicit, time-limited product feature, but the consent prompt does not clearly disclose that transcript content is sent to the model provider and defaults to yes. Warn about this privacy exposure; the source does not establish an npm install-time attack.

Static reason

High-risk behavior combination matched malicious policy.

Trigger

A user runs interactive setup and accepts or leaves blank the onboarding-log consent prompt; later, the staging and analysis scripts run while that consent remains active.

Impact

Session content may be processed by the configured model provider despite the prompt describing only a distilled summary as shared. The behavior is consent-gated and time-limited.

Mechanism

The staging script reads completed local transcripts. The analyzer places transcript turns in a prompt sent to the Claude CLI, then posts the resulting friction summary to the GDS inbox.

AI rationale

Evidence

scripts/gds/setup.jsscripts/gds/onboarding-ship.jsscripts/gds/onboarding-analyze.jsmodules/onboarding/onboarding.jsmodules/grading/grader-subagent.js

Decision evidence

public snapshot

AI called this Suspicious at 84.0% confidence as Dangerous Capability with medium false-positive risk.

Evidence for warning

The manifest defines server and CLI entrypoints but no npm lifecycle script, so this behavior is not triggered by package installation.
Interactive setup asks to analyze onboarding logs, says raw transcripts stay local, and accepts a blank response as consent; noninteractive setup records no consent.
With consent active, the onboarding scripts locate local session transcripts and stage older files; the analyzer builds a prompt from transcript turns and sends it to the configured Claude CLI.
The analyzer posts the resulting friction summary to the GDS inbox. The source shows transcript content is also sent to the model for analysis, creating a privacy exposure that the setup prompt does not clearly disclose.
The agent-integration targets resolve to CLI materialization, templates, worktree detection, or protected paths. The cited skill hook only reads edited skill paths and prints lint warnings; no account-changing action was identified.

Evidence against

The onboarding flow requires interactive setup consent and stops staging when consent is absent or its time window has expired.
The setup prompt says only a distilled summary is shared, and noninteractive setup does not record consent.

Source & flagged code

23 flagged

tests/public_mirror_export.mjsView file

Lines 83-123javascript

83 assert.equal(isBinary(Buffer.from('plain text file\nwith lines')), false);

84 assert.equal(isBinary(Buffer.from('')), false);

85});

87// ── redactBlob ───────────────────────────────────────────────────────────────────────

88t('redactBlob redacts a text blob and reports zero leaks', () => {

89 const r = redactBlob({ buf: Buffer.from('serve https://widgetco.com to zoltan at 8.8.8.8'), rules, sensitiveLiterals, envSecrets: [] });

90 const text = r.buf.toString('utf8');

91 assert.ok(!/widgetco|zoltan|8\.8\.8\.8/i.test(text), text);

92 assert.deepEqual(r.leaks, []);

93 assert.equal(r.secret, false);

94});

95t('redactBlob keeps a binary blob verbatim but still leak-scans it', () => {

96 const bin = Buffer.concat([Buffer.from('PNG\x00'), Buffer.from('widgetco'), Buffer.from('\x00')]);

97 const r = redactBlob({ buf: bin, rules, sensitiveLiterals, envSecrets: [] });

98 assert.ok(r.buf.equals(bin), 'binary must be byte-identical (not corrupted by text redaction)');

99 assert.ok(r.leaks.length > 0, 'a literal embedded in a binary must still trip the gate');

100});

101t('redactBlob SCRUBS a secret shape from a text blob (handled, not leaked)', () => {

102 const r = redactBlob({ buf: Buffer.from('key AKIA' + 'ABCDEFGHIJKLMNOP here'), rules, sensitiveLiterals, envSecrets: [] });

103 assert.ok(!r.buf.toString('utf8').includes('AKIAABCDEFGHIJKLMNOP'), 'the secret must be scrubbed from the published text');

Critical

Critical Secret

Package contains a critical-looking secret pattern.

tests/public_mirror_export.mjsView on unpkg · L103

104 assert.equal(r.secret, false, 'a scrubbed text secret is handled, so it is not a residual leak');

105});

106// Regression (task 1003873). The secret verdict audits THIS call's own scrub, so it has to

107// share the scrub's knownDocNames allowlist. When it did not, the audit re-redacted the ADR

108// filename the scrub had just spared, read the difference as a leak, and failed the packager

109// closed on every doc in the tree — the publish lane stayed red for seven core releases.

110t('redactBlob reports no secret for an allowlisted doc filename (the gate agrees with its own scrub)', () => {

111 const name = '0281-differential-audit-allowlist-symmetry.md';

112 const knownDocNames = new Set([name]);

113 const r = redactBlob({ buf: Buffer.from(`see [ADR 0281](docs/adr/${name}) for why`), rules, sensitiveLiterals, envSecrets: [], knownDocNames });

114 assert.ok(r.buf.toString('utf8').includes(name), 'the allowlisted filename must survive the scrub');

115 assert.equal(r.secret, false, 'and must not then be reported as a residual secret');

116 assert.deepEqual(r.leaks, []);

117});

118t('redactBlob FAILS CLOSED on a secret inside a BINARY blob (cannot safely rewrite it)', () => {

119 const bin = Buffer.concat([Buffer.from('BIN\x00'), Buffer.from('AKIA' + 'ABCDEFGHIJKLMNOP')]);

120 const r = redactBlob({ buf: bin, rules, sensitiveLiterals, envSecrets: [] });

121 assert.ok(r.buf.equals(bin), 'binary kept verbatim');

122 assert.equal(r.secret, true, 'a secret in a binary must trip the gate (the export then publishes nothing)');

123});

103patternName = aws_access_key severity = critical line = 103

Critical

Secret Pattern

AWS access key ID in tests/public_mirror_export.mjs

tests/public_mirror_export.mjsView on unpkg · L103

bin/bongos.jsView file

Lines 1-39javascript

1#!/usr/bin/env node

2// bin/bongos.js — the Cloud Bongos CLI dispatcher.

3//

4// GDS-V4 · V4.R79 · task 1246 (Phase 1 of the `bongos` toolkit; epic 1245).

5//

6// A single `bongos <verb>` front door over the existing scripts/gds/* scripts —

7// the same actions the /builder-* slash-commands run inside Claude Code, now

8// usable from any terminal, VS Code, a Chromebook, or CI. This is deliberately a

9// THIN dispatcher: it spawns the target script with `process.execPath` (the one

10// guaranteed cross-platform runtime — no bash, #686) and forwards argv + exit

scripts/setup-builder.jsView file

71// and never spread from caller-supplied opts: shell:true on Windows runs the L72: // args through cmd.exe expansion, so it's used ONLY for `npm` (a .cmd shim that L73: // can't be spawned without a shell) and ONLY with hardcoded args — never with

High

Shell

Package source references shell execution.

scripts/setup-builder.jsView on unpkg · L71

public-docs/redoc.standalone.jsView file

8patternName = generic_password severity = medium line = 8

Medium

Secret Pattern

Hardcoded password in public-docs/redoc.standalone.js

public-docs/redoc.standalone.jsView on unpkg · L8

1/*! For license information please see redoc.standalone.js.LICENSE.txt */ L2: !function(e,t){"object"==typeof exports&&"object"==typeof module?module.exports=t(require("null")):"function"==typeof define&&define.amd?define(["null"],t):"object"==typeof exports... L3: || (${a} == "string" && ${i} && ${i} == +${i})`).assign(l,s._`+${i}`);case"integer":return void n.elseIf(s._`${a} === "boolean" || ${i} === null

Low

Eval

Package source references a known benign dynamic code generation pattern.

public-docs/redoc.standalone.jsView on unpkg · L1

tests/hall_profile_world.mjsView file

Lines 18-58javascript

18// ledgers are one key over, so the first screen is a screen.

19// * THE CHARTS ARE ARITHMETIC FIRST. cumulativeByDay folds two ships on one

20// day into one point and accumulates, over whichever timestamp the caller

21// names; craftMix folds the tail past seven into "other" rather than

22// minting an eighth hue; weeklyBuckets is a time axis (every week present,

23// empty or not).

24// * DRACHMAE COME FROM THE LEDGER (task 1003491). The earned line is the

25// credits ledger — /me/credits on your own profile, /builders/:id/credits

26// on another's — `delta` cumulated by `recorded_at`, so it ends at the

scripts/gds/fitness-ratchets.jsView file

Lines 19-59javascript

19// duplicate_window_count — normalized WINDOW-line blocks appearing in 2+

20// distinct files (the in-house jscpd: cross-file copy/paste)

21// require_cycle_count — strongly-connected components (size >1) in the

22// relative-require graph (the in-house madge)

23// console_call_count — console.* call sites in SERVER paths (src/ + modules/

24// minus their browser public/ trees); task 1003130 converts the core path

25// to structured logging and shrinks this baseline

26// doc_stale_links / doc_broken_anchors / doc_duplicate_prose — the

27// docs-entropy scanner's three counts, reused as a library

modules/ui-design/kit/serve.jsView file

Lines 7-47javascript

7// module may not require core and the doorway may not import the serving domain,

8// so the real code path is reached as a child process) — plus the content-hash

9// ?v= stamp, with the instance's own branding pack. Answers the public feeds from

10// fixtures/ so a page boots with real-shaped data and no database. Serves docs/

11// raw so a mock under docs/design/mocks/ renders with its relative links.

12//

13// node modules/ui-design/kit/serve.js -> http://127.0.0.1:3412/ (the public-landing module)

14// SURFACE=<dir or module key> a checkout dir ("modules/hall-ui/public") or a module

15// key ("public-landing") whose manifest declares a webSurfaces dir

scripts/gds/build-bongos.jsView file

24const path = require('node:path'); L25: const { spawnSync } = require('node:child_process'); L26: const { arg } = require('./cli-lib'); L27: L28: const REPO_ROOT = path.join(__dirname, '..', '..'); L29: const ENTRY = path.join(REPO_ROOT, 'bin', 'bongos.js'); ... L42: function bunPresent() { L43: const r = spawnSync('bun', ['--version'], { encoding: 'utf8', shell: process.platform === 'win32' }); L44: return !!r && !r.error && r.status === 0; ... L66: console.error('✖ `bun` is required to compile the standalone binary but was not found.'); L67: console.error(' Install it (https://bun.sh) — e.g. `curl -fsSL https://bun.sh/install | bash` —'); L68: console.error(' or in CI use the oven-sh/setup-bun action. Then re-run.');

High

Sandbox Evasion Gated Capability

Source gates dangerous network, credential, or execution behavior behind CI, host, platform, time, or geo fingerprint checks.

scripts/gds/build-bongos.jsView on unpkg · L24

modules/security/artifact-scan/fetch.jsView file

10L11: const dns = require('node:dns'); L12: const fsp = require('node:fs/promises'); L13: const path = require('node:path'); L14: const { spawn } = require('node:child_process'); L15: ... L19: const MAX_CAPTURE_BYTES = 1 * 1024 * 1024; L20: // git's stderr can name internal hosts and paths; only a short excerpt travels, L21: // and only in a finding's detail, never in a message a caller might echo. ... L50: if (a === 0) return 'the unspecified range'; L51: if (a === 10) return 'a private (RFC 1918) address'; L52: if (a === 172 && b >= 16 && b <= 31) return 'a private (RFC 1918) address';

High

Cloud Metadata Access

Source reaches cloud instance metadata or link-local credential endpoints.

modules/security/artifact-scan/fetch.jsView on unpkg · L10

tests/grader_prompt_fencing.mjsView file

74contains invisible/control Unicode U+202E (right-to-left override) const dirty = `a${'<U+202E>'}b${'<U+202D>'}c${'<U+202A>'}d`;

Critical

Trojan Source Unicode

Source contains bidi control or invisible Unicode characters associated with Trojan Source attacks.

tests/grader_prompt_fencing.mjsView on unpkg · L74

docker-entrypoint.shView file

•path = docker-entrypoint.sh kind = build_helper sizeBytes = 1075 magicHex = [redacted]

Medium

Ships Build Helper

Package ships non-JavaScript build or shell helper files.

docker-entrypoint.shView on unpkg

modules/autonomy/pollers/routine-timers.js#virtual:normalized:round1View file

•stage = ast_semantic_analysis; reason = ast_parse_error; limitedFiles = 84

High

Semantic Analysis Limited

A bounded semantic-analysis stage reached its safety limit; remaining detectors completed, but this package requires AI review.

modules/autonomy/pollers/routine-timers.js#virtual:normalized:round1View on unpkg

tests/audit_middleware.mjsView file

365patternName = generic_password severity = medium line = 365

Medium

Secret Pattern

Hardcoded password in tests/audit_middleware.mjs

tests/audit_middleware.mjsView on unpkg · L365

tests/provision.mjsView file

3053patternName = generic_password severity = medium line = 3053

Medium

Secret Pattern

Hardcoded password in tests/provision.mjs

tests/provision.mjsView on unpkg · L3053

3091patternName = generic_password severity = medium line = 3091

Medium

Secret Pattern

Hardcoded password in tests/provision.mjs

tests/provision.mjsView on unpkg · L3091

tests/auto_merge_diagnosis.mjsView file

297patternName = github_pat severity = critical line = 297

Critical

Secret Pattern

GitHub personal access token in tests/auto_merge_diagnosis.mjs

tests/auto_merge_diagnosis.mjsView on unpkg · L297

tests/logger.mjsView file

62patternName = github_pat severity = critical line = 62

Critical

Secret Pattern

GitHub personal access token in tests/logger.mjs

tests/logger.mjsView on unpkg · L62

tests/secret_gate.mjsView file

21patternName = github_pat severity = critical line = 21

Critical

Secret Pattern

GitHub personal access token in tests/secret_gate.mjs

tests/secret_gate.mjsView on unpkg · L21

tests/key-status.mjsView file

25patternName = google_api_key severity = high line = 25

High

Secret Pattern

Google API key in tests/key-status.mjs

tests/key-status.mjsView on unpkg · L25

Source & flagged code

23 flagged · loading source

tests/public_mirror_export.mjsView file

103patternName = aws_access_key severity = critical line = 103

Critical

Critical Secret

Package contains a critical-looking secret pattern.

tests/public_mirror_export.mjsView on unpkg · L103

103patternName = aws_access_key severity = critical line = 103

Critical

Secret Pattern

AWS access key ID in tests/public_mirror_export.mjs

tests/public_mirror_export.mjsView on unpkg · L103

bin/bongos.jsView file

19const path = require('node:path'); L20: const { spawnSync } = require('node:child_process'); L21:

High

Child Process

Package source references child process execution.

bin/bongos.jsView on unpkg · L19

Findings

6 Critical8 High10 Medium7 Low

CriticalCritical Secrettests/public_mirror_export.mjs

CriticalTrojan Source Unicodetests/grader_prompt_fencing.mjs

CriticalSecret Patterntests/public_mirror_export.mjs

CriticalSecret Patterntests/auto_merge_diagnosis.mjs

CriticalSecret Patterntests/logger.mjs

CriticalSecret Patterntests/secret_gate.mjs

HighChild Processbin/bongos.js

HighShellscripts/setup-builder.js

HighSame File Env Network Executionmodules/ui-design/kit/serve.js

HighSandbox Evasion Gated Capabilityscripts/gds/build-bongos.js

HighCloud Metadata Accessmodules/security/artifact-scan/fetch.js

Affected versions and remediation

This report applies to @bongos/core@1.20.50.

See all versions of @bongos/core for other recorded verdicts.

Review the evidence and your use of @bongos/core@1.20.50 before allowing it. Restrict the permissions described in this report.
Choose an independently verified alternative or release. This report does not establish that other versions are safe.

Evidence last updated: 2026-10-18 01:04:21.599Z (UTC).

Think this verdict is wrong? Dispute it on GitHub or read how reports are produced.

AI Security Review

Decision evidence

Source & flagged code

Source & flagged code

Findings

Affected versions and remediation

Behavioral surface

Source & flagged code