Meta paper: Agents complete tasks but fail to prevent catastrophic actions like factory resets
rohanpaul_ai · x · 2026-09-02
A new Meta paper introduces ADeptS-Bench, highlighting a critical failure in computer-use agents: they lack consequence reasoning. All 7 tested models proceeded with a $25K checkout and failed to identify a button labeled "Optimize" that actually triggered a factory reset.
Ablation studies show that removing explicit refusal tools significantly increased attack success rates for top models (Gemini, Claude, GPT). The findings suggest that current "agent safety" relies heavily on wrapper logic, not just the model itself.
More from coding & agent
- LLM one-shots a Minecraft 3x3 piston door by building its own Linux test infra — Vjeux · 2026-10-18
- Dev builds Mumbai-inspired 3D manga background generator from OpenStreetMap data — CurieuxExplorer · 2026-10-18
- Solo founder builds a money-printing data service on Claude's Pi Durable harness — MicahBerkley · 2026-10-18
- Dev builds Arch-based OS with custom desktop and browser using Claude Code — Snoo-10493 · 2026-10-18
- Five SaaS agent workflows fail the same way — and it's never model intelligence — usehive · 2026-10-18
- Claude skill generates hand-drawn interactive diagrams that teach kids how things work — CatAstro_Piyush · 2026-10-18