Five red teamers, same AI tools, five wildly different results
rez0__ · x · 2026-11-16
HackingLZ observes that giving five offensive-security people of varying skill levels the same Codex or Claude access and the identical codebase or network target yields five very different results.
rez0 mostly agrees but finds it confusing: his own steering feels minimal—small prompt differences, skill file tweaks, and subagent orchestration choices. He suspects these small differences add up to explain the huge variance in outcomes.
More from coding & agent
- KubeCon's big theme is inference: a full-stack breakdown from DRA GPU slicing to vLLM and llm-d — AccBalanced · 2026-11-16
- Swyx on AI Engineer's three themes: software factories, System 1 models, ambition — thursdai_pod · 2026-11-16
- Hiker replaced a $60/yr app with a self-built clone: 12 minutes, $0.62 in API costs via Hermes Agent — Teknium · 2026-11-16
- Expose local MCP servers to ChatGPT securely via OpenAI's official tunnel — no auth or public URL needed — altryne · 2026-11-16
- Dev replicates DeepSeek's web search UI and exposes what its queries actually look like — teortaxesTex · 2026-11-16
- Every built a writing skill from 70 articles so Claude Fable 5.2 writes for their columnist — every · 2026-11-16