The AI review rates the findings as likely false positive, but the risk score (100/100) still counts them.
Analysis record
- Analysed
- 1 month ago 1 October 2026
- Version
- v2.0.2
- Artifact
- SHA256 b6a29a8a…05259d
- Detail
- Partial list
Is @opichi/mcp-api-helper safe?
@opichi/mcp-api-helper is an MCP server that plugs Perplexity AI into an agent workflow so the assistant can help with API integrations. This package declares no special permissions, and the endpoint list it ships carries no domain tied to the publisher. What it does carry is a full Python virtual environment under python/venv, and that folder is the source of nearly every flag in this scan.
The critical obfuscation alerts all point at compiled .so files such as python/venv/lib/python3.12/site-packages/black/linegen.cpython-312-darwin.so. Those are the machine-code builds of Black, a widely used Python formatter, and pydantic_core, a data validation library. Compiled binaries read as opaque to any scanner, so a file that cannot be parsed as text gets marked as hiding something. The seven MCP-TRANSPORT-HARDCODED-TOKEN findings sit in starlette/datastructures.py, pydantic/types.py and httpx/_urls.py, three general purpose libraries where token-shaped constants show up in ordinary parsing code.
The network list is the clearest tell. Entries like activate.fish and 64-linux-gnu.so are filenames from inside that virtual environment, scraped as though they were web addresses. None of them resolve to a site this server would contact. No tool description carries instructions aimed at an AI model either, which is the failure that would actually matter for an MCP server.
Nothing in this package points to behaviour beyond what it advertises. A scanner read a bundled Python environment and flagged the parts it could not parse.
Evidence ledger
Ranked by severity · findings with a source location link to the code viewer
YARA Rule Matches
14 rules| Severity | Rule | Hits | Files | Metadata |
|---|---|---|---|---|
| LOW | DebuggerStatementsShouldNotBeUsed | 27 | python/venv/lib/python3.12/site-packages/pygments/lexers/__pycache__/lisp.cpython-312.pycpython/venv/lib/python3.12/site-packages/dotenv/__pycache__/main.cpython-312.pycpython/venv/lib/python3.12/site-packages/pygments/lexers/_asy_builtins.py +24 more | - |
| LOW | UsingCommandLineArguments | 1 | bin/mcp-api-helper.js | - |
| LOW | spyeye | 6 | python/venv/lib/python3.12/site-packages/pip/_vendor/distlib/wheel.pypython/venv/lib/python3.12/site-packages/uvicorn/protocols/http/__pycache__/httptools_impl.cpython-312.pycpython/venv/lib/python3.12/site-packages/cryptography/hazmat/primitives/kdf/__pycache__/kbkdf.cpython-312.pyc +3 more | - |
| LOW | credential env files | 121 | python/venv/lib/python3.12/site-packages/pip/_vendor/distro/distro.pyREADME.mdpython/venv/lib/python3.12/site-packages/markdown_it/rules_inline/image.py +118 more | - |
| LOW | RedirectToUnknownPath | 8 | python/venv/lib/python3.12/site-packages/authlib/integrations/flask_client/apps.pypython/venv/lib/python3.12/site-packages/httpx/_client.pypython/venv/lib/python3.12/site-packages/pip/_vendor/requests/auth.py +5 more | - |
| LOW | postinstall persistence mechanism | 48 | python/venv/lib/python3.12/site-packages/uvicorn/_types.pypython/venv/lib/python3.12/site-packages/uvicorn/main.pypython/venv/lib/python3.12/site-packages/uvicorn/supervisors/basereload.py +45 more | - |
| LOW | WeakSSLTLSProtocolsShouldNotBeUsed | 7 | python/venv/lib/python3.12/site-packages/pip/_vendor/urllib3/contrib/_securetransport/low_level.pypython/venv/lib/python3.12/site-packages/pip/_vendor/urllib3/contrib/_securetransport/__pycache__/low_level.cpython-312.pycpython/venv/lib/python3.12/site-packages/pip/_vendor/urllib3/connection.py +4 more | - |
| LOW | NoDisableSanitizeHtml | 1 | python/venv/lib/python3.12/site-packages/markdown_it_py-3.0.0.dist-info/METADATA | - |
| LOW | postinstall file download | 266 | python/venv/lib/python3.12/site-packages/authlib/oauth2/rfc9101/authorization_server.pypython/venv/lib/python3.12/site-packages/authlib/integrations/base_client/sync_openid.pypython/venv/lib/python3.12/site-packages/pygments/lexers/scripting.py +263 more | - |
| LOW | NoUseEval | 36 | python/venv/lib/python3.12/site-packages/pydantic/_internal/_typing_extra.pypython/venv/lib/python3.12/site-packages/pip/_vendor/rich/markup.pypython/venv/lib/python3.12/site-packages/cffi/recompiler.py +33 more | - |
| LOW | postinstall environment access | 152 | python/venv/lib/python3.12/site-packages/certifi-2025.6.15.dist-info/METADATApython/venv/lib/python3.12/site-packages/pydantic/v1/version.pypython/venv/lib/python3.12/site-packages/mcp/client/sse.py +149 more | - |
| LOW | postinstall obfuscation | 184 | python/venv/lib/python3.12/site-packages/pip/_internal/operations/install/wheel.pypython/venv/lib/python3.12/site-packages/pygments/lexers/configs.pypython/venv/lib/python3.12/site-packages/httpcore/_sync/connection.py +181 more | - |
| LOW | postinstall registry modification | 73 | python/venv/lib/python3.12/site-packages/authlib/oauth2/rfc6749/authenticate_client.pypython/venv/lib/python3.12/site-packages/pygments/lexers/text.pypython/venv/lib/python3.12/site-packages/dns/rdtypes/dsbase.py +70 more | - |
| LOW | postinstall file manipulation | 25 | python/venv/lib/python3.12/site-packages/pip/_internal/cli/main_parser.pypython/venv/lib/python3.12/site-packages/pip-25.1.1.dist-info/RECORDpython/venv/lib/python3.12/site-packages/pygments/lexers/foxpro.py +22 more | - |
Finding Categories
YARA Rules Matched
14 rules(955 hits)MCP Server Analysis
MCP servers expose tools and resources to AI assistants. Unlike browser extensions, they run as standalone processes with direct system access. Tool definitions are analyzed for prompt injection, data exfiltration, and tool poisoning patterns.
AI Security Report
AI Security Review
Evidence context: threat category none; evidence quality moderate.
Tool poisoning
The scan matched no tool-poisoning patterns at all. That is the finding that matters most for an MCP server, because the threat defining this package class is a tool description carrying instructions aimed at the model. There is nothing of the sort here: no directives telling an agent to stay quiet, no invisible Unicode runes in tool names, no XML-style instruction tags in metadata. The package presents itself as an API integration helper built on Perplexity AI, and the surface it exposes matches that description.
What the critical findings actually are
All 34 obfuscation alerts carry the title OBFUSCATION-supply_chain_binary, and every one lands on a compiled extension inside the bundled virtual environment: python/venv/lib/python3.12/site-packages/black/linegen.cpython-312-darwin.so, black/parsing.cpython-312-darwin.so, black/mode.cpython-312-darwin.so and pydantic_core/_pydantic_core.cpython-312-darwin.so. These are mypyc-compiled builds of Black and pydantic_core, two of the most commonly installed packages in the Python world. A .so file cannot be read as source, and the rule treats unreadable binaries as suspicious. That mechanism accounts for every critical obfuscation hit.
The seven secret findings are titled MCP-TRANSPORT-HARDCODED-TOKEN and sit in python/venv/lib/python3.12/site-packages/starlette/datastructures.py, pydantic/types.py, httpx/_urls.py, cryptography/hazmat/_oid.py and authlib/oauth1/rfc5849/signature.py. Five separate general purpose libraries, each shipping constants that look token-shaped to a pattern matcher. A genuine hardcoded transport token would live in this server's own transport configuration. These live inside an OAuth signature helper.
Credentials and network
Nothing in the findings reads .ssh, .aws/credentials, .kube/config or application_default_credentials.json. The package touches no sensitive paths. The endpoint list is the clearest artefact of automated extraction: it contains abi3.so, activate.fish, 64-linux-gnu.so, 1ctypesbackend.new and a0--0.ac, which are filenames and hex fragments pulled from inside the virtualenv and treated as web addresses. Mixed in with them are real infrastructure entries nobody would call exfiltration, including acm.org and example.com. The single network finding sits in the same bundled code as everything else.
The count of 4586 indicators and 3441 code-smell items follows from the same cause. A virtual environment with hundreds of vendored libraries produces thousands of matches, and none of them describe behaviour this package author wrote.
The counterargument
The strongest case against a clean verdict is the distribution shape: an MCP server that ships an entire Python virtual environment is asking for trust, and a future release could swap one of those .so files and turn this into a real supply-chain problem. That concern is about version 3.0.0, not about the code in front of us. Version 2.0.2 contains no hidden instructions, no credential harvesting and no calls to domains outside Perplexity's documented API. Nothing here separates this package from any other Python project scanned the same way.
Key Reasons
- Zero tool-poisoning findings: no hidden AI directives, invisible Unicode, or instruction tags in any tool metadata.
- All 34 obfuscation alerts are OBFUSCATION-supply_chain_binary on compiled .so files in python/venv (black, pydantic_core), which is normal for mypyc-built Python packages.
- The 7 MCP-TRANSPORT-HARDCODED-TOKEN hits land in third-party libraries (starlette/datastructures.py, pydantic/types.py, httpx/_urls.py, authlib OAuth signature code), not in the server's own transport config.
- No credential file targeting: nothing touches .ssh, .aws, .kube or application_default_credentials.json.
- The network endpoint list is IoC extractor noise containing filenames and fragments such as activate.fish, 64-linux-gnu.so and abi3.so rather than resolvable domains.
- No package permissions or host permissions are declared, and no tool description carries behavior aimed at a model.
False Positive Considerations
- Bundled Python venv with compiled native extensions triggering binary/obfuscation rules
- Generic hardcoded-token pattern matching against constants inside library source files
- IoC extractor scraping filenames and hex fragments as domain names
- Code-smell rules firing across thousands of bundled dependency files
Reviewed 2026-10-02 · version 2.0.2; recommended action: suppress false positive; model confidence 80%.
Source Code Not Available
Source code is not available for this version of the extension.
About This Extension
Frequently Asked Questions
Similar Extensions
Related extensions from the same publisher or marketplace