registry / Data exfiltration / @ooples/token-optimizer-mcp / 7.5.0

@ooples/token-optimizer-mcp@7.5.0

Audit net MCP context avoided per AI coding agent, optimize context, and share a live local knowledge graph across 16 CLI clients

AI Security Review

scanned 9d ago · by lpm-firewall-ai

The package does not activate agent hooks during an ordinary install, but its enabled runtime harvester sends session prompts and conclusions remotely by default when an API key is available. This is a real conditional disclosure risk, so warn rather than block.

Static reason

High-risk behavior combination matched malicious policy.; source matched previously finalized malicious package; routed for review

Trigger

An edited agent session reaches its Stop hook while remote harvesting is enabled by an available API key and the user has not opted out.

Impact

Session information can leave the machine without a separate affirmative opt-in to remote harvesting; the code provides an opt-out setting.

Mechanism

The hook passes the transcript to a worker that builds a digest and posts it to the configured model endpoint. The digest includes prompts, conclusions, paths, and commands, but omits tool results and file contents.

AI rationale

Evidence

hooks-core/stop-harvest.mjshooks-core/harvest.mjspackage.jsonscripts/postinstall.cjs

Network endpoints1

api.anthropic.com/v1/messages

Decision evidence

public snapshot

AI called this Suspicious at 90.0% confidence as Dangerous Capability with medium false-positive risk.

Evidence for warning

The postinstall hook is declared in the manifest, but it installs hooks only when the auto-install environment flag is set; a normal global install reports that nothing changed and directs the user to run the installer.
The Stop hook starts harvesting only after session edits and a debounce check, then passes the transcript to a detached worker.
Harvesting uses an existing API key for remote mode unless the user opts out; the default remote endpoint is Anthropic’s Messages API.
The digest can include user prompts, assistant conclusions, file paths, and commands, while tool results are skipped.
The harvester sends the digest in a network request to its configured endpoint and attaches the matching API credential when present.
Other inspected integration locations concern client hook destinations, settings or configuration lookup, plugin diagnosis, uninstall gating, and an API fetch helper; they did not establish a separate authenticated account action.

Evidence against

The package does not install agent hooks on an ordinary package install; automatic setup requires an explicit environment flag.
The harvest can be disabled with its opt-out setting, and the inspected digest omits tool results and file contents.

Source & flagged code

14 flagged

package.jsonView file

Lines 42-82json

42 "!integrations/**/lib/**",

43 "plugin/**/*",

44 "scripts/**/*",

45 "!**/logs/**",

46 "!**/*.log",

47 "!**/.cache/**",

48 "!**/.token-optimizer/**",

49 "!**/*.db",

50 "!**/*.db-*",

51 "!**/*.sqlite",

52 "!**/*.sqlite-*",

53 "!**/*.sqlite3",

54 "!**/*.sqlite3-*",

55 "!**/*.tmp",

56 "!**/*.temp",

57 "!**/.token-optimizer-edit-*.lock",

58 "!**/.env",

59 "!**/.env.*"

60 ],

61 "scripts": {

62 "postinstall": "node scripts/postinstall.cjs",

High

Install Time Lifecycle Scripts

Package defines install-time lifecycle scripts.

package.jsonView on unpkg · L62

Medium

Ambiguous Install Lifecycle Script

Install-time lifecycle script is not statically allowlisted and needs review.

package.jsonView on unpkg · L62

63 "checksums": "node scripts/checksums.mjs",

64 "screenshots": "node scripts/capture-screenshots.mjs",

65 "uninstall-hooks": "node scripts/uninstall.mjs",

66 "doctor": "node scripts/doctor.mjs",

67 "learn": "node dist/learn/cli.js",

68 "diagnostics": "node scripts/export-diagnostics.mjs",

69 "projects:discover": "node scripts/discover-project-graphs.mjs",

70 "build": "tsc && npm run copy:assets",

71 "copy:assets": "node scripts/copy-assets.mjs",

72 "start": "node dist/server/index.js",

73 "dev": "tsc --watch",

74 "clean": "rm -rf dist",

75 "prepublishOnly": "npm run clean && npm run build",

76 "test": "node --experimental-vm-modules node_modules/jest/bin/jest.js",

77 "test:watch": "node --experimental-vm-modules node_modules/jest/bin/jest.js --watch",

78 "test:ci": "node --experimental-vm-modules node_modules/jest/bin/jest.js --ci --coverage --maxWorkers=2",

79 "test:coverage": "node --experimental-vm-modules node_modules/jest/bin/jest.js --coverage",

80 "test:integration": "node --experimental-vm-modules node_modules/jest/bin/jest.js tests/integration",

81 "test:unit": "node --experimental-vm-modules node_modules/jest/bin/jest.js tests/unit",

82 "test:benchmark": "node --experimental-vm-modules node_modules/jest/bin/jest.js tests/benchmarks",

hooks-core/skeleton.mjsView file

Lines 9-49javascript

9 * can be strictly BETTER than the file:

10 *

11 * auth.ts -- 91 KB, structure + what we know

12 * function

hooks-core/harvest.mjs#virtual:normalized:round1View file

255* Calls the model. Returns [] on any failure -- a harvest that errors must be L256: * indistinguishable, from the caller"s side, from a session with nothing to\n * learn. `harvestFailure()` carries why, for the diagnostics that do care.\n */\n/**\n * The findings ... L257: * joining argv with single spaces and no quoting at all, which breaks the

High

Shell

Package source references shell execution.

hooks-core/harvest.mjs#virtual:normalized:round1View on unpkg · L255

295// the prompt-file route exists to keep off a Windows command line: L296: // cmd.exe cannot carry a newline inside an argument, and the quoting L297: // doubles every `"`. These CLIs document the flag value as being\n // APPENDED to stdin, so the whole payload rides stdin and the argument\n // is one line of ASCII that any...

High

Command Output Exfiltration

Source combines command execution, command-output handling, and outbound requests; review data flow before blocking.

hooks-core/harvest.mjs#virtual:normalized:round1View on unpkg · L295

295Trigger-reachable command-output exfiltration chain: manifest.bin -> scripts/doctor.mjs -> hooks-core/doctor.mjs -> hooks-core/harvest.mjs L295: // the prompt-file route exists to keep off a Windows command line: L296: // cmd.exe cannot carry a newline inside an argument, and the quoting L297: // doubles every `"`. These CLIs document the flag value as being\n // APPENDED to stdin, so the whole payload rides stdin and the argument\n // is one line of ASCII that any...

dist/compress/onnx.jsView file

Lines 41-81javascript

41 return hash % vocabSize;

42 });

43 };

44}

45/**

46 * Loads a model and returns an encoder over it.

47 *

48 * Throws if `onnxruntime-node` is absent or the model will not load, because

49 * this is called once at startup by a caller who explicitly asked for a model.

dist/tools/configuration/smart-env.jsView file

Lines 110-150javascript

110 // Analyze variables

111 const result = await this.analyzeEnvironment(parsed, content, options, filePath, fileHash);

112 // Cache result

113 await this.

hooks-core/harvest.mjsView file

Lines 4-44javascript

4 * Runs at Stop/PreCompact against a cheap model, in a separate process, so the

5 * session doing the work never pays for the harvest. Without an API key the

6 * whole thing is skipped and the graph keeps its structural layer, which is the

7 * majority of its machinery and costs nothing.

plugin/hooks/lib/restore.mjsView file

3contains invisible/control Unicode U+FEFF (zero width no-break space) <U+FEFF>/**

Critical

Trojan Source Unicode

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

plugin/hooks/lib/restore.mjsView on unpkg · L3

install-hooks.ps1View file

•path = install-hooks.ps1 kind = build_helper sizeBytes = 21337 magicHex = [redacted]

Medium

Ships Build Helper

Package ships non-JavaScript build or shell helper files.

install-hooks.ps1View on unpkg

dist/analysis/project-analyzer.jsView file

•stage = ast_semantic_analysis; reason = ast_alias_growth_limit_exceeded; limitedFiles = 31

High

Semantic Analysis Limited

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

dist/analysis/project-analyzer.jsView on unpkg

cli-wrapper.mjsView file

•matchType = normalized_sha256 matchedPackage = @ooples/token-optimizer-mcp@7.1.0 matchedPath = cli-wrapper.mjs matchedIdentity = npm:[redacted]:7.1.0 similarity = 1.000 summary = normalized source hash matched finalized malicious source

High

Known Malware Source Similarity

Source file is highly similar to a previously finalized malicious package; route for source-aware review.

cli-wrapper.mjsView on unpkg

scripts/managed-policy.mjsView file

•matchType = normalized_sha256 matchedPackage = @ooples/token-optimizer-mcp@7.1.0 matchedPath = scripts/managed-policy.mjs matchedIdentity = npm:[redacted]:7.1.0 similarity = 1.000 summary = normalized source hash matched finalized malicious source

High

Known Malware Source Similarity

Source file is highly similar to a previously finalized malicious package; route for source-aware review.

scripts/managed-policy.mjsView on unpkg

Source & flagged code

14 flagged · loading source

package.jsonView file

•scripts.postinstall = node scripts/postinstall.cjs

High

Install Time Lifecycle Scripts

Package defines install-time lifecycle scripts.

package.jsonView on unpkg

•scripts.postinstall = node scripts/postinstall.cjs

Medium

Ambiguous Install Lifecycle Script

Install-time lifecycle script is not statically allowlisted and needs review.

package.jsonView on unpkg

hooks-core/skeleton.mjsView file

29L30: import { execFileSync } from 'node:child_process'; L31: import { dirname } from 'node:path';

High

Child Process

Package source references child process execution.

hooks-core/skeleton.mjsView on unpkg · L29

Findings

1 Critical9 High6 Medium7 Low

CriticalTrojan Source Unicodeplugin/hooks/lib/restore.mjs

HighInstall Time Lifecycle Scriptspackage.json

HighChild Processhooks-core/skeleton.mjs

HighShellhooks-core/harvest.mjs#virtual:normalized:round1

HighSame File Env Network Executionhooks-core/harvest.mjs

HighCommand Output Exfiltrationhooks-core/harvest.mjs#virtual:normalized:round1

HighTrigger Reachable Command Output Exfiltrationhooks-core/harvest.mjs#virtual:normalized:round1

HighSemantic Analysis Limiteddist/analysis/project-analyzer.js

HighKnown Malware Source Similaritycli-wrapper.mjs

HighKnown Malware Source Similarityscripts/managed-policy.mjs

Affected versions and remediation

This report applies to @ooples/token-optimizer-mcp@7.5.0.

See all versions of @ooples/token-optimizer-mcp for other recorded verdicts.

Review the evidence and your use of @ooples/token-optimizer-mcp@7.5.0 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: 2026-10-14 07:41:58.164Z (UTC).

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

@ooples/token-optimizer-mcp@7.5.0

AI Security Review

Decision evidence

Source & flagged code

Source & flagged code

Findings

Affected versions and remediation

Runtime surface

Behavioral surface

Source & flagged code