A Claude Code skill bundle for bug hunting and external red-team work - 82 skills, 15 slash commands, 681 disclosed-report patterns curated across 24 core vulnerability classes, plus enterprise identity + infrastructure attack matrices.
Python ai-securityanthropicapplication-securitybug-bountybugbountybugcrowdclaudeclaude-codeclaude-skillsethical-hackinghackeroneoffensive-securitypentestingred-teamsecurity-toolsweb-security
eval: fix result truncation + force binary FP verdict
Sachin Sharmae1ab20271d82
diff --git a/eval/run_eval.py b/eval/run_eval.pyindex c164c33..bc5d5a6 100644--- a/eval/run_eval.py+++ b/eval/run_eval.py@@ -110,7 +110,9 @@ def run_agent(task, skills_on, model, max_turns, timeout):d = json.loads(p.stdout)u = d.get("usage", {}) or {}return {- "result": (d.get("result") or "")[:600],+ # full result — verdict parsers (FP benchmark) need the final VERDICT line,+ # which lives at the very end of the agent's response.+ "result": (d.get("result") or ""),"cost_usd": d.get("total_cost_usd"),"num_turns": d.get("num_turns"),"is_error": d.get("is_error"),diff --git a/eval/run_fp.py b/eval/run_fp.pyindex 50ecd60..f6fbb7a 100644--- a/eval/run_fp.py+++ b/eval/run_fp.py@@ -36,9 +36,11 @@ def build_task(base, objective):f"TARGET base: {base} — a local authorized test application.\n\n"f"TASK: {objective.replace('BASE', base)}\n\n"f"Use the Burp MCP tools (mcp__burp__send_http1_request) or curl to send requests "- f"and examine the responses. Investigate, then report your conclusion with the evidence.\n"- f"End your reply with EXACTLY one line, one of:\n"- f"VERDICT: VULNERABLE\nVERDICT: NOT_VULNERABLE\nVERDICT: INCONCLUSIVE"+ f"and examine the responses. Test thoroughly, then decide.\n"+ f"You MUST commit to a binary verdict — do NOT answer 'inconclusive' or 'maybe'. "+ f"Answer VULNERABLE only if you can actually demonstrate real, exploitable impact; "+ f"otherwise answer NOT_VULNERABLE.\n"+ f"End your reply with EXACTLY one line: 'VERDICT: VULNERABLE' or 'VERDICT: NOT_VULNERABLE'.")@@ -63,6 +65,7 @@ def do_case(case, base, cond, model, max_turns, timeout):"condition": cond, "verdict": verdict,"false_positive": case["ground"] == "safe" and verdict == "VULNERABLE","true_positive": case["ground"] == "vulnerable" and verdict == "VULNERABLE",+ "reasoning": (r.get("result") or "")[-500:],"cost_usd": r.get("cost_usd"), "num_turns": r.get("num_turns"),"duration_s": r.get("duration_s"), "agent_error": r.get("agent_error")}