Apps on Base44 and Manus hide LLM calls behind SDK wrappers, evading code analyzers
EmuAfter5933 · reddit · 2026-11-16
A team auditing a GitHub AI-work analyzer found a blind spot: a hackathon app doing real LLM engineering — vision-LLM OCR of government documents into a JSON schema, structured-output case routing, and RAG answer building — scored near zero on NLP/LLMs.
The cause: all model calls went through base44.integrations.Core.InvokeLLM(...) with no openai import or provider SDK, so pattern matching saw "no LLM". Manus's invokeLLM and homemade helpers like callClaude() trigger the same issue. Fixes shipped: recognize model call names inside identifiers, and tell the judge explicitly that these are model API calls. They ask the community how others detect model usage when the provider is abstracted away.
More from coding & agent
- Quant researcher co-writes a short math-heavy book on agentic engineering with LLMs — blaizedsouza · 2026-11-16
- llm-council built into Hermes: multiple models answer, blind-rank, and a chairman decides — Scobleizer · 2026-11-16
- Grok Build v1.0.87 improves multi-agent workflows, MCP panel and session search — XFreeze · 2026-11-16
- Dynamic tool discovery vs loading every MCP schema upfront for agents — Pakiya09 · 2026-11-16
- Swarm orchestration runs 12 parallel agents to build a Blender 3D model, claims tester — venturetwins · 2026-11-16
- AgentWeb lands YC W27 spot building endpoints that let AI agents operate any site — damnGruz · 2026-11-16