Caper Pre-Executes Shell Commands in Sandboxes for Safer Agent Approvals
SingaporeManagementUniversity · hf · 2026-11-04
SMU researchers introduce Caper, a consequence-aware approval mechanism for autonomous coding agents. Unlike text-based approvers, Caper pre-executes each shell command in a snapshot sandbox, gives an independent approver a bounded report of observed effects, commits allowed commands, and rolls back blocked ones with actionable reasons. Compared with six existing systems, Caper achieves the lowest harmful-outcome rate on SABER (6.7% vs at least 14.8%).
More from coding & agent
- With: an MIT-licensed systems language repo with 4,200+ commits and agentic dev workflow built in — QuixiAI · 2026-11-04
- GitHub team shares ~30 min of Copilot demos, from editor to Slack workflows — lee_stott · 2026-11-04
- Microsoft's WebHarbor docks 153 real websites into dockerized environments for GUI agent RL — microsoft · 2026-11-04
- ToolSpace mines real MCP execution traces to synthesize tool environments for long-horizon agents — RUC · 2026-11-04
- Turn count misleads agent benchmarks: DEPTH measures sequential dependency depth instead — Yu Jiang · 2026-11-04
- Karpathy builds HackerTube in ~2 hours with zero code using Opus 5.5 — FinanceYF5 · 2026-11-04