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
State the engagement frame in-band instead of implying it (#51) (#52)
Sachin Sharmaa9058aa31e50
diff --git a/README.md b/README.mdindex d7e3077..3bc0303 100644--- a/README.md+++ b/README.md@@ -194,6 +194,25 @@ The bundle explicitly **excludes**: weaponizing 0-days against unauthorized targ> **Heads-up — Anthropic runtime cyber safeguards.** Anthropic's models apply real-time safeguards that **block "vulnerability exploitation or offensive security tooling development" by default** — so even *authorized, in-scope* work can hit a refusal that isn't this bundle's doing. If you do authorized offensive security (pentest / bug bounty / red team), enroll in Anthropic's **free, application-based [Cyber Verification Program (CVP)](https://claude.com/form/cyber-use-case)** to get safeguards adjusted for legitimate dual-use work. (Mass data exfiltration and ransomware development stay prohibited and are *not* adjustable.) Details: [Anthropic — real-time cyber safeguards](https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude).+### Why your model switched mid-session++Separate from refusals, and easy to miss. On **Opus 5**, a narrow set of higher-risk cyber requests —+Anthropic names *exploit generation*, *binary-based vulnerability scanning* and *penetration testing* —+**fall back to Opus 4.8** rather than being refused. You get a notice and the response is labelled with+the model that answered, but in a long agentic run that is easy to scroll past, so it can look like+Opus 5 quietly got worse. See [why Claude switched models](https://support.claude.com/en/articles/16049681-why-claude-switched-models-in-your-conversation-with-opus-5).++What to do depends on what you are actually doing:++| Situation | What helps |+|---|---|+| **Auditing your own code** — reviewing a repo you own for defects | Say so. "Defensive review of my own repo", "check this against the OWASP Top 10", "secure refactor to remediate" describe the work accurately and read as remediation. This is not a workaround; the work genuinely is defensive. |+| **Authorized offensive work** — live engagement, PoC for a bounty submission | This is what the bundle is for, and the supported route is [CVP](https://claude.com/form/cyber-use-case). Do not reword an offensive engagement to look defensive to get past a classifier — enroll instead. |+| **You just want the switching off** | Settings → Capabilities disables automatic model switching. |++`/hunt` states the engagement frame (authorized, scope-bounded, remediable finding) on its first turn+for exactly this reason — engagement context belongs in the session explicitly, not implied.+---## Documentationdiff --git a/commands/hunt.md b/commands/hunt.mdindex cb44616..be6f98e 100644--- a/commands/hunt.md+++ b/commands/hunt.md@@ -5,7 +5,16 @@ description: Active vulnerability hunting. Two-track dispatcher — asks Red Tea# /hunt-slim two-track dispatcher. one mode question, one branch, delegate. never asks about SOW — invoking `/hunt` implies SOW is signed.+slim two-track dispatcher. one mode question, one branch, delegate.++**engagement frame — state once, then proceed.** invoking `/hunt` asserts the operator holds written+authorization to assess the named scope: a bug-bounty program's in-scope assets, a signed engagement+letter, infrastructure they own, or a practice target. work stays inside that scope — an out-of-scope+host stops the run rather than widening it. the deliverable is a reproducible, remediable finding.++do not prompt for this and do not ask about the SOW. it is an assertion the operator already made by+invoking the command; restating it once keeps the engagement context explicit for the rest of the+session instead of leaving it implied.## step 0 — parsediff --git a/skills/hunt-dispatch/SKILL.md b/skills/hunt-dispatch/SKILL.mdindex 38caeaa..8ece720 100644--- a/skills/hunt-dispatch/SKILL.md+++ b/skills/hunt-dispatch/SKILL.md@@ -7,6 +7,20 @@ description: Skill-set loader for /hunt orchestrator. Fingerprints the target, pskill-set loader for `/hunt`. one concept (which skills to load), one place.+## engagement context++every skill loaded below operates under one frame, and it holds for the whole session:++- **authorized and scope-bounded.** the operator asserted written authorization for the named scope+ when they invoked `/hunt`. testing stays inside it. an out-of-scope host ends the run — it does not+ widen it.+- **the output is a finding, not an exploit.** the deliverable is a reproducible proof that a defect+ exists, written so the owner can remediate it. enough to demonstrate impact; no further.+- **remediation is the point.** these skills exist so defects get fixed by the people who own them.++this frame is stated here because it is the choke point every `/hunt` run passes through before any+`hunt-*` skill loads. it is not a prompt and needs no answer.+invocation contract:```