Testing agents for prompt injection: why cloud models keep flagging your red-team runs
cocktailpartyfixer · reddit · 2026-11-19
An engineer who red-teams his company's agents shares his workflow: a coding agent writes injection attacks, fires them at the target agent, and judges the replies — all wrapped into a skill.
The pain point is cloud models: the same injection passes one day and gets flagged the next, sometimes mid-run, since providers treat injections hidden in emails as abuse even against his own staging environment. He wants a local model (64 GB Mac) for the attacker side — convincing injections, no refusals, coding skill irrelevant — and asks about abliterated models or finetunes people have actually used. He also asks what sources people follow for current agent-defense research, noting most material he finds dates to 2024 while defenses have moved on.
More from coding & agent
- nohumans.directory bought 8,000+ paid x402 APIs with real USDC so your agent knows before it pays — SashSail · 2026-11-19
- MemSkill: Memory Skills Learned with LLaMA Transfer to Qwen Without Retraining — jiqizhixin · 2026-11-19
- Honcho Launches Dream Histories to Inspect How Its Memory System 'Dreams' About Users — Scobleizer · 2026-11-19
- Six-year PKM coach: a markdown dump folder isn't a second brain — build a schema.md for your agent — evielync · 2026-11-19
- FrontierRSI benchmark: Claude Opus 5.5 leads at 34.3%, GPT-6 models trail by ~10 points — nathanbenaich · 2026-11-19
- RL-tuned Qwen3.8-27B for opencode hits Mac via MLX at 156 tok/s on M5 Max — yb2698 · 2026-11-19