feat(benchmarks): add CyBench evaluation — ZAK + GPT-5 reaches 9/11 (…

[infosecsanket]

…81.8%) on official strict-offline subset

Phase 1 (curated 9-task offline slice from HackTheBox Cyber Apocalypse 2024):

• ZAK + gpt-5: 9/9 (100.0%) pass@2 at $2.99
• ZAK + gpt-4o-2024-11-20 on identical slice: 4/9 (44.4%) at $1.34
• 55 percentage-point same-task model-uplift gap with harness, prompt, scoring, and iteration cap held constant

Phase 2 (11-task strict-offline subset of cybench's official task_list.txt, spanning Very-Easy through Hard difficulty):

• ZAK + gpt-5 unguided pass@2: 8/11 (72.7%)
• ZAK + gpt-5/o3 best-mode pass@2: 9/11 (81.8%) (subtask-guided + selective o3 escalation on Partial Tenacity)
• Includes Hard-tier Permuted crypto solve

To our knowledge this is the first published GPT-5 result on cybench-format CTF tasks; no GPT-5 entry currently exists on the public leaderboard at cybench.github.io.

Includes:

• run_cybench.py — single-file harness, ~700 LoC, sole dep openai>=2.28
• PAPER.pdf — 17-page academic-style writeup
• SCORECARD.md / REPORT.md / REPORT.html — multi-audience summaries
• LEADERBOARD_PR.md — submission package for cybench.github.io
• 48 per-task transcript JSONs (every LLM turn + every shell command) organized by configuration: unguided / subtask-guided / o3-escalation / phase1-curated / gpt4o-baseline

Total project cost: $34. Reproducible end-to-end.