Security audit
Security checks for vulnerabilities and agentic risk
This skill is a coherent marketing-copy scorer, but users should understand that normal scoring can send pasted copy to Anthropic if the local backend is unavailable or Haiku is selected.
Install only if you are comfortable with pasted marketing copy being processed by a local LLM server or by Anthropic Claude when Haiku is used or local auto-detection fails. Use --demo or --compliance-only for no remote call, set LLM_BACKEND=local for sensitive drafts, and treat the TRIBE v2/fMRI calibration claims as unverified unless the publisher provides the referenced calibration artifacts.
The public description understates actual behavior: the skill performs remote LLM analysis, hook comparison, and compliance scanning beyond the narrow scoring description. Description-behavior mismatches are dangerous because they defeat informed consent, interfere with security review, and can cause sensitive copy or internal business text to be sent to third-party services unexpectedly.
The advertised calibration suggests a concrete scoring methodology or model lineage tied to TRIBE v2 and fMRI brain-response data. In code, scoring is produced by prompting either a local OpenAI-compatible model or Anthropic Haiku and combining outputs with hard-coded weights; there is no dataset, model artifact, calibration routine, or reference to TRIBE v2 beyond the manifest claim.
The skill invokes networked and environment-backed capabilities via Anthropic API access but does not declare any explicit tool scope or permissions. That creates a transparency and governance gap: users and orchestration systems may treat the skill as lower-risk than it is, while pasted content can still be transmitted externally.
The skill invites users to paste marketing copy for premium scoring without a clear warning that the content will be transmitted to Anthropic. If users submit unreleased campaigns, customer data, or proprietary sales material, this can produce unintended third-party disclosure and policy noncompliance.
The file repeatedly endorses and rewards 'embedded commands,' 'assume-the-close,' urgency, and other NLP persuasion tactics in natural-language instructions and examples. While not a static security issue, this is a semantic policy concern because the skill encourages covert influence techniques as a default behavior rather than transparently framing them as optional or constrained.
The manifest says the skill scores marketing copy for resonance, hook strength, NLP technique usage, and conversion readiness, returning a 0-100 score with breakdown and rewrite suggestions. The code also exposes a forbidden-words compliance checker, a standalone hook comparison mode, and a full rewrite generator as first-class CLI capabilities, which materially expand the skill beyond the described scoring scope.
The script description emphasizes local MLX scoring by default, but the implementation can fall back to Anthropic and transmit user-provided marketing copy off-host. This creates a confidentiality and consent risk because users may supply proprietary or regulated content believing processing is local-only.
User copy is sent to the Claude backend without a runtime warning or consent prompt when auto-fallback occurs. If users input confidential campaign drafts, client data, or sensitive business text, the skill may silently exfiltrate that content to a third-party service.
Code issues a request to a loopback, link-local, or private-range host. This can reach internal services not meant to be exposed and is a common SSRF pivot.
try:
import openai
import urllib.request
urllib.request.urlopen("http://localhost:8800/health", timeout=1)
return openai.OpenAI(base_url="http://localhost:8800/v1", api_key="local"), "local"
except Exception:
import anthropic
The manifest markets the skill as a calibrated scoring engine, but actual premium behavior depends on a third-party hosted LLM and an API key. This is a trust and disclosure issue: users may assume deterministic local analysis when their content is actually processed by Anthropic, creating unexpected confidentiality and compliance exposure.
The manifest and main description present this as a scoring/rewrite skill operating on provided copy, with implementation centered on score_content.py. However, the documentation explicitly directs recalibration via vault/learnings/2026-03-27-tribe-v2-colab-spec-task47.md, indicating dependence on or reference to external project data beyond the described scoring workflow.
No suspicious patterns detected.