Why it exists
A security skill's value is judgment, not a script. Point an agent at a target and the hard part is not running a tool, it is knowing what to look at first, in what order, and when a lead is real. These skills write that discipline down: enumerate the whole taxonomy before you fixate on one family, treat rank as triage and never a filter, and hold that a lead is a fact to adjudicate, never a verdict.
Because that reasoning is the value, every skill here is tool-agnostic. Each one names the capability a step needs, for example "something that answers who calls this from a real parse", without prescribing a product. Bring your own tools. The method transfers to anyone, whatever they run.
What you get
Method you can load, not a scanner you configure.
Method, not a script
Each skill encodes the reasoning, ordering, and adjudication of real black-box and white-box testing, the part a scanner cannot give you.
Tool-agnostic by design
A skill names the capability a step needs, not a product. Bring a code graph, a static analyzer, or careful manual tracing. The loop is the same.
Thirty-three skills, six lanes
White-box, research, AI-agent, black-box, cloud, and reporting, from guard-gap audits to prompt-injection tests to IAM privilege-escalation paths.
Discipline baked in
Enumerate the whole taxonomy, rank is triage not a filter, a lead is a fact not a verdict. The habits that separate a real hunt from a demo.
One shared finding schema
Every skill emits findings in the same schema, so results from any lane are consistent, deduplicable, and ready to write up without reformatting.
Authorized use only
Every skill opens with a scope check or gate. Nothing here targets systems you do not have permission to test. Your own code, OSS, CTFs, and in-scope work.
Install
Add the marketplace and install once, and every skill becomes available to your agent, matched by name and description. Or copy a single skill by hand.
/plugin marketplace add UnboundCompute/security-agent-skills /plugin install security-agent-skills@unboundcompute # all skills, updatable cp -r skills/detecting-race-conditions .claude/skills/ # or one skill, by hand
Not on Claude Code? Each SKILL.md is a standalone playbook. Point any agent runtime that follows the Agent Skills format at the skills/ directory, or read one and follow the loop by hand.
The six lanes
Coverage runs from source-level bug hunting to agent and cloud attack surface, with reporting to close the loop.
white-box guard gaps, taint adjudication, memory safety, races, business logic, crypto, authz research bug variants, n-day from a patch, dependency-CVE reachability, supply chain agent lethal trifecta, prompt injection, MCP tool abuse, multi-agent, RAG poisoning, guardrails black-box attack-surface recon, cache attacks, SAML and OIDC, request smuggling, DOM, SSRF cloud IAM privilege escalation, CI/CD OIDC trust, non-human identity reachability reporting a confirmed finding to a reproducible writeup
The white-box lane assumes you have some way to answer structural questions from a real parse, a code property graph, a static analyzer, or careful tracing on a small target. That is exactly what Lachesis provides, though the method does not depend on which tool you bring.
These skills encode the same testing discipline our hosted autonomous security researcher runs at scale. Read them, run them yourself, or let the researcher do it for you.
See the researcher →