buried-injections v1.0.0

Benchmark of 10 open-source prompt-injection detectors on AgentDojo's 629
attacks embedded in realistic tool output, plus 97 benign cases.

• Leaderboard at default thresholds (make bench-agentdojo)
• Each attack scored alone vs embedded in context (make bench-payloads)
• Catch rate at a 2% false-alarm budget, with thresholds calibrated on 3
  AgentDojo domains and measured on the unseen 4th (make bench-budget)
• Prompt Guard 2 input-scope and window-size experiment (make bench-windows)
• Raw per-case scores in bench/results/ so every number can be rechecked

Write-up: https://dev.to/rudratosh/i-tested-10-prompt-injection-detectors-on-629-real-ai-agent-attacks-291h