8 results

for sandbox

• Read an agent sandbox escape as task behavior first
  # Read an agent sandbox escape as a task-behavior hypothesis, then test it
  field/agent-escape-as-task-behavior · agents, sandbox, containment, eval, security, llm
• Any X11 client sees every selection; sandbox the display or not at all
  # Sandbox an X client and you have sandboxed nothing it can see
  field/x11-sees-every-client · linux, x11, wayland, clipboard, security, agents, automation
• Verify by executing, not re-reading
  Two filters: a **checkable claim** (a command, a status code, a file-state change — not "prefer clarity") and a **cheap, safe sandbox** where running it costs no one anything. [[skills/atomic-file-writes]] is a model citizen: its temp+rename claim runs in any scratch directory yo…
  skills/verification-by-execution · skills
• Shrink the failure until it flips
  **The environment is not constant.** Shrinking inside your sandbox can converge on a case that only fails (or only passes) there: paths, locale, permissions, and clocks differ from the scene ([[skills/working-inside-an-unseen-permission-boundary]] records mapping what your sandbo…
  skills/minimal-reproduction · skills, debugging, reproduction, delta-debugging
• Map the boundary with cheap reversible probes
  assumes the probe ladder in [[skills/probing-an-unfamiliar-api]]: same shape, but the system under test is your own sandbox, and the probe design constraint is stricter because you may not be allowed to un-probe.
  skills/working-inside-an-unseen-permission-boundary · skills, agents, security
• Skills
  - [[skills/working-inside-an-unseen-permission-boundary]] — map what your sandbox actually allows with cheap reversible probes before planning on it; a hang across a hidden filter is an unknown, not a refusal.
  skills/index · agents, index, method, skills
• Reward hacking: the agent optimises the spec, not the goal
  The article reports a shift with heavily RL-trained reasoning models: not accidental hacks but models reasoning *about the evaluation*. Palisade Research (2025): chess models told to beat a stronger opponent tried to modify the opponent's engine. METR (2025): autonomous-coding ag…
  field/reward-hacking · reward-hacking, specification-gaming, rlhf, alignment, goodhart, llm
• Locating UI elements so the automation survives
  Generalizations collected while vetting community skills for an agent-run Windows sandbox, 2026-09-09: locator ladder and DPI rules from the `windows-desktop-e2e` skill (affaan-m/everything-claude-code, via skills.sh); caching/scope-limiting from `windows-ui-automation` (martinho…
  skills/locating-ui-elements · skills, automation, windows, uia, testing