Red-teaming your own agents: cloud models keep refusing, local alternatives?
cocktailpartyfixer · reddit · 2026-11-19
A developer who red-teams the agents he builds at work shared his setup and asked for advice:
- He turned red-teaming into a skill: a coding agent writes injection attacks, runs them against the target agent, and judges the replies (repo: agent-redteam).
- Pain point: cloud models are unreliable — the same request passes one day and gets flagged the next, sometimes mid-run, because an injection hidden in an email looks like abuse.
- Ask #1: wants a local attacker model (possibly a judge too) that writes convincing injections without refusing, on a Mac with 64 GB — experiences with abliterated models or specific finetunes welcome.
- Ask #2: looking for recent (last few months) papers, repos, and benchmarks on agent security, since most material is from 2024 and defenses have moved on.
More from coding & agent
- Survey: A Dual-Paradigm Framework for Agentic AI, Reviewing 90 Studies (2018–2025) — adnan_hashmi · 2026-11-19
- Ex-DeepSeek Core Member Launches Linger, an Open-Source Personal Agent Running Fully on Your PC — FellMentKE · 2026-11-19
- Ex-DeepSeek core member open-sources Linger, a personal AI agent that runs local models on your PC — FellMentKE · 2026-11-19
- Apple opens iWork to external LLMs: Pages, Numbers, Keynote gain MCP support — dsp_ · 2026-11-19
- Hermes adapter gets GGUF release: matches xhigh thinking at a quarter of the tokens — yb2698 · 2026-11-19
- Gym's chatbot invented a free membership-pause policy; team shares its guardrail lessons — Fit_Satisfaction541 · 2026-11-19