Security risk report for @forgecat/andrej-karpathy-skills v0.0.8
Source Integrity
SRC-010Medium
Profile is a converted/republished artifact from upstream GitHub repository (multica-ai/andrej-karpathy-skills), introducing supply-chain dependency on external source.
ForgeCat conversion process adds metadata layer (README.md, file references) but content originates from third-party maintainer; no cryptographic verification or pinning of upstream commit visible in profile metadata.
Details
Findings
Data exfiltration
EXAMPLES.md · SKILL-002
Agent Intent
Low
Content is educational guidance on coding best practices (simplicity, clarity, test-driven development) with no instructions to manipulate the AI, read credentials, exfiltrate data, or bypass safety mechanisms.
Examples demonstrate *what LLMs do wrong* (overcomplication, hidden assumptions) and *how to fix it* (surface assumptions, simplify, verify)—this is descriptive documentation of pitfalls, not prescriptive injection.
No system prompt leakage, guidance poisoning, or role-hijacking language detected; the four principles (Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution) are legitimate software engineering practices that improve code quality and transparency.
Details
Evidence
'State your assumptions explicitly. If uncertain, ask.' — encourages clarity, not deception.
'If you write 200 lines and it could be 50, rewrite it.' — promotes simplicity, not obfuscation.
'Every changed line should trace directly to the user's request.' — enforces accountability, not hidden behavior.
Examples section explicitly labels ❌ (wrong) vs ✅ (correct) patterns; no malicious instructions embedded.
Permissions
Low
Profile declares no tools, hooks, or MCP servers; it is pure markdown guidance text.
No file_write, file_delete, shell, web_fetch, or other high/medium-risk tool categories requested or configured.
Content is read-only instructional material intended to influence agent reasoning, not grant executable authority.
Details
Evidence
'Declared tools (normalized categories; paths are untrusted): (none declared)'
'MCP servers (untrusted data): (none)'
MCP Risk
Low
No MCP servers defined in the profile.
No tool descriptions, binary execution, network access, or filesystem scope to evaluate.
ForgeCat