Security testing skills
Security testing skills are curated knowledge packages that help Strix agents move from broad security knowledge to focused testing. They support attack-surface discovery, vulnerability validation, technology-aware analysis, and scan strategies that match the time and coverage needs of an assessment.
Security knowledge model
The skill library gives an agent focused guidance for the part of a security assessment it is performing. The knowledge describes useful attack paths, test methods, and ways to confirm a real issue so that the agent can connect reconnaissance to evidence rather than treat every signal as a finding.
Testing knowledge domains
The library groups knowledge by the context that changes how a test should be planned. Vulnerability skills explain a weakness class. Reconnaissance and tooling skills help build and inspect the attack surface. Analysis skills help agents challenge candidates and record what remains unproved. Framework, protocol, technology, cloud, and custom skills explain the behavior of a particular environment.
| Domain | Testing purpose |
|---|---|
| Vulnerabilities | Explore common and advanced weakness classes and validate their impact. |
| Reconnaissance | Discover assets, entry points, and changes in the external attack surface. |
| Tooling | Use security tools in a controlled way and interpret their results. Opengrep is the current source-triage path for bundled offline rules. Browser guidance supports safe page interaction and traffic-aware web testing. |
| Analysis | Test whether a candidate is real, safe, or still unproved. |
| Frameworks and protocols | Adapt tests to application structure and communication rules. Current guidance spans Django, FastAPI, NestJS, Next.js, GraphQL, and OAuth. |
| Technologies and cloud | Test service-specific trust boundaries, identity rules, and configuration risks across AWS, Kubernetes, Active Directory, Auth0, observability systems, LLM applications, and Supabase. |
| Custom guidance | Add focused workflows for repository or industry-specific needs. |
The library also includes coordination and source-aware guidance for assessments that have repository access. The source-aware white-box workflow combines Opengrep triage, structural AST mapping, syntax-aware parsing, secret checks, and dependency or configuration checks before dynamic validation. This keeps static output as a source of test hypotheses rather than a substitute for proof. The skill library and loading page describes the complete organization and lookup rules.
These domains work together. For example, asset discovery can reveal an API, a protocol skill can explain its request model, and a vulnerability skill can guide authorization tests against the discovered resources. The same pairing applies when a cloud or platform service creates a new trust boundary, or when an application framework changes how input and identity reach a sink.
References: strix/…/README.md• L28-52 strix/…/source_aware_whitebox.md• L8-35 strix/…/source_aware_sast.md• L18-64 strix/…/__init__.py• L58-81 strix/…/aws.md• L1-8 strix/…/kubernetes.md• L1-8 strix/…/django.md• L1-8 strix/…/fastapi.md• L1-8 strix/…/nestjs.md• L1-8 strix/…/nextjs.md• L1-8 strix/…/graphql.md• L1-8 strix/…/oauth.md• L1-8 strix/…/active_directory.md• L1-8 strix/…/auth0.md• L1-8 strix/…/grafana_prometheus.md• L1-8 strix/…/llm_applications.md• L1-12 strix/…/supabase.md• L1-8 strix/…/agent_browser.md• L8-20
Skill guidance
Each package adds expert guidance instead of acting as a separate scanner. It can turn a broad task into a test plan by defining useful inputs, target-specific probes, practical variations, and the evidence needed to confirm a real issue. The guidance covers both established application weaknesses and newer boundaries such as browser state, agent authority, LLM data paths, and observability pivots. Workflow skills also explain how to run a validated pentest or place a focused, diff-scoped check in CI without changing the evidence standard.
Browser guidance adds a controlled way to inspect and exercise web applications. It helps an agent use page structure, isolated sessions, saved authentication state, network capture, and safe recovery rules without treating page content as instructions. This makes browser interaction useful for reconnaissance and proof while preserving the assessment's authorization and evidence boundaries.
Source-aware guidance connects code signals to runtime behavior. The white-box coordination skill expects a source map, a static triage pass, and a structural pass before specialists use those results to choose dynamic probes. The source-aware SAST skill makes Opengrep the default first pass and adds AST, syntax, secret, dependency, and JavaScript checks when the target needs them. It also directs agents to inspect cross-component meaning and namespace fallback when a local pattern does not explain the full risk.
Technology guidance keeps that reasoning specific to the target. Django guidance, for example, joins URL and API surfaces with ORM, middleware, session, serializer, admin, upload, and channel boundaries. It also requires version and deployment evidence for behavior that depends on a framework release or a configured component. This lets a specialist test the deployed trust boundary rather than apply a generic framework checklist.
Security guidance can also separate related outcomes that need different proof. The XSS skill treats context, transformation order, and the final sink as the basis for testing. It distinguishes HTML injection from script execution and follows stored values through later render paths, so a blocked script does not hide another concrete browser effect and an unproved execution claim does not become a false finding.
Focused vulnerability guidance adds similar proof rules for different input and trust models. IDOR guidance binds the subject, action, and object across accounts and transports. NoSQL guidance tests operator and structure injection, while query-language guidance treats each scalar search or authorization field as an independent sink and requires paired broadening and narrowing evidence. Dependency CVE guidance is a source-scan path for known advisories: it preserves findings that do not need a dynamic exploit and records package, advisory, lockfile, and reachability evidence through the dependency report path.
Semantic-confusion guidance covers cases where several components assign different meanings to the same attacker-influenced value. It has the agent model the transformation path, identify where a security decision occurs, and compare that representation with the one consumed by the final sink. This is useful for parser differentials, normalization drift, overloaded fields, lifecycle state, protocol translation, and namespace fallback.
Protocol and state-change guidance makes cross-system trust rules explicit. GraphQL guidance connects schema discovery to resolver, transport, federation, subscription, and input checks, including exact validation for newer schema constraints. OAuth guidance follows the authorize-to-token chain and tests redirect binding, session state, proof keys, issuer, audience, and account-linking effects. CSRF guidance checks whether those state changes remain protected across methods, content types, transports, and downstream consumers.
For example, API guidance uses schemas and credentials with different rights to support endpoint discovery and authorization tests. General pentest guidance connects scan goals to authorized targets and validated proof. It supports both a self-hosted CLI and a managed cloud path on the same testing engine, so teams can choose local control or shared service features without changing the finding model. CI guidance applies the same scan engine to pull-request changes and keeps the result usable as a merge gate through findings, exit status, and completion state.
Counterevidence guidance gives every candidate a clear closure path. An agent marks a candidate confirmed only with a working proof or a complete reachable source-to-impact trace. It marks a candidate ruled_out only when a named control blocks the attacker's path before the dangerous effect. If proof and a named control are both missing, it records open_proof_gap instead of treating uncertainty as safety. Negative controls and an honest confidence level help distinguish a working defense from an unreachable or broken test.
Source-aware discovery guidance keeps code-based assessment tied to concrete reachability. It keeps separate candidates for independently reachable instances and distinct impacts, preserves both the entrypoint and the shared sink, and directs family sweeps across common trust-boundary and parser patterns. This prevents a safe sibling or a shared helper from hiding another reachable weakness.
The same guidance keeps closure visible in the coverage record. Reported, ruled-out, follow-up, no-issue, and not-applicable outcomes show whether a surface was tested and what the result means. Agents can revise a prior result when new reachability or control evidence changes the assessment, while the history preserves the reason for the change.
This makes a skill useful at the point where a general agent needs a focused answer. The agent can use the guidance to choose the next probe, understand a response, or decide what evidence is needed before it reports a finding.
References: strix/…/README.md• L60-72 strix/…/source_aware_whitebox.md• L14-47 strix/…/source_aware_sast.md• L66-140 strix/…/dependency_cve_scanning.md• L8-23 strix/…/dependency_cve_scanning.md• L115-152 strix/…/django.md• L8-35 strix/…/django.md• L157-215 strix/…/opengrep.md• L11-66 strix/…/agent_browser.md• L26-80 strix/…/agent_browser.md• L407-425 strix/…/agent_browser.md• L533-540 skills/…/SKILL.md• L10-61 skills/…/SKILL.md• L10-35 skills/…/SKILL.md• L55-110 skills/…/SKILL.md• L156-162 skills/…/SKILL.md• L10-17 skills/…/SKILL.md• L21-23 skills/…/SKILL.md• L68-72 strix/…/counterevidence.md• L8-39 strix/…/counterevidence.md• L100-143 strix/…/counterevidence.md• L145-184 strix/…/source_aware_discovery.md• L7-15 strix/…/source_aware_discovery.md• L17-42 strix/…/source_aware_discovery.md• L65-89 strix/…/idor.md• L172-195 strix/…/nosql_injection.md• L241-267 strix/…/query_language_injection.md• L10-36 strix/…/query_language_injection.md• L113-145 strix/…/semantic_confusion.md• L6-44 strix/…/semantic_confusion.md• L46-109 strix/…/semantic_confusion.md• L165-198 strix/…/xss.md• L1-8 strix/…/xss.md• L47-78 strix/…/xss.md• L186-211 strix/…/graphql.md• L138-140 strix/…/graphql.md• L264-281 strix/…/oauth.md• L78-118 strix/…/oauth.md• L129-149 strix/…/csrf.md• L81-86 strix/…/csrf.md• L151-175 strix/…/llm_applications.md• L1-12
Assessment strategies
Scan-mode skills set the overall balance between speed, coverage, and depth. The general pentest guidance exposes quick, standard, and deep choices for an assessment, while dedicated mode guidance defines the detailed behavior for each strategy. CI guidance applies the same strategy model to change-focused checks. These skills do not replace vulnerability knowledge. They tell the agent how much reconnaissance to perform, which targets to prioritize, and how far to follow an attack path.
Scan modes
Strix provides broad assessment strategies for time-boxed, balanced, deep, and change-focused work:
| Mode | Primary aim | Testing shape |
|---|---|---|
| Quick | Find high-impact issues fast. | Time-boxed checks of critical flows with limited discovery and focused proof. Source-backed work starts with changed or high-risk paths and a lightweight structural pass. |
| Standard | Balance coverage and effort. | Systematic attack-surface review followed by structured testing and validation. Source-backed work uses first-pass Opengrep triage and bounded structural mapping. |
| Deep | Maximize coverage and attack-path depth. | Broad reconnaissance, detailed workflow analysis, and extensive chaining. Source-backed work combines broad static and structural triage with detailed manual review. |
| Diff | Review a supplied change set. | Follow changed behavior far enough to find new or newly reachable security problems while excluding unrelated old issues. |
The choice affects what the agent considers sufficient work. Quick mode favors fast confirmation of serious issues. Standard mode provides a repeatable assessment across the main attack surfaces. Deep mode spends more time mapping the system and pursuing multi-step impact. Diff mode narrows the reportable scope without lowering the evidence bar. It anchors findings to the changed behavior and records coverage by changed component. In CI, that scope must resolve to the actual pull request base so that all changed commits remain covered.
The source-aware modes share a common evidence rule. Opengrep, AST tools, secret scanners, and dependency checks help rank paths and expose hypotheses, but static output does not become a vulnerability report until the relevant behavior is validated. Quick mode keeps that bundle narrow. Standard and deep modes expand it as coverage and attack-path depth require.
The agent tools and workflows page explains how these strategies fit into the wider tool and agent workflow.
References: skills/…/SKILL.md• L55-110 skills/…/SKILL.md• L21-23 skills/…/SKILL.md• L68-72 skills/…/SKILL.md• L86-110 strix/…/quick.md• L8-24 strix/…/standard.md• L8-25 strix/…/deep.md• L8-34 strix/…/source_aware_sast.md• L18-64 strix/…/opengrep.md• L33-66 strix/…/diff.md• L7-35 strix/…/diff.md• L38-57 strix/…/diff.md• L62-87
Context-specific testing
An assessment can combine a scan mode with focused knowledge for the target. A web application may need a framework skill and an authorization skill. An API may need protocol guidance and a vulnerability skill. For example, GraphQL testing can pair schema and transport guidance with authorization or CSRF checks, while OAuth testing can pair flow guidance with CSRF and account-linking checks. A cloud or platform review may need service guidance paired with identity, input, or access-control testing. A source-backed assessment can also add source-aware guidance so the agent can connect code paths to runtime behavior.
The system keeps this combination focused. A specialist normally receives a small set of related skills, while broader work stays with the parent agent or uses more than one specialist. This lets the assessment cover different risk areas in parallel without giving every agent the full library.
The coordinating root agent turns this focused knowledge into a shared assessment plan. It derives a threat model for the scan, delegates target work to specialists, and reconciles open coverage items before completion. In black-box work, reconnaissance supplies the first model. Specialists can amend the shared model as they learn more, so later tests use the same view of attackers, assets, and trust boundaries.
References: strix/…/prompt.py• L20-57 strix/…/tools.py• L415-453 strix/…/root_agent.md• L8-37 strix/…/root_agent.md• L39-43 strix/…/source_aware_whitebox.md• L37-47 strix/…/aws.md• L157-186 strix/…/kubernetes.md• L176-195 strix/…/nextjs.md• L216-235 strix/…/graphql.md• L263-280 strix/…/oauth.md• L134-149 strix/…/csrf.md• L158-175 strix/…/auth0.md• L147-163 strix/…/supabase.md• L252-275 strix/…/agentic_system_security.md• L146-155 strix/…/browser_security.md• L155-174
Skill loading
Skill knowledge enters an agent in two ways. The scan prompt can preload guidance that shapes the agent from the start. During a scan, the agent can request additional reference material when it reaches a technology or workflow that was not known in advance.
The runtime searches registered extra skill roots before the built-in skill root. This lets a deployment add new guidance or override packaged guidance without editing the package. The most recently registered extra root has the highest precedence, so a matching skill path resolves to that root.
Prompt context
Prompt context is the agent's baseline security knowledge for its current role. It includes the requested specialist guidance, the selected scan strategy, and shared operating guidance. Root and source-aware agents can receive extra coordination guidance when their role requires it.
This context gives the agent a stable testing frame before it uses tools. It also keeps mode guidance present throughout the scan while allowing the requested specialist knowledge to remain targeted.
References: strix/…/prompt.py• L20-57 strix/…/prompt.py• L60-110
On-demand reference
On-demand loading is a short-lived way to bring exact skill guidance into the current conversation. An agent uses it when it needs a specific test method or payload pattern that was not preloaded. The added material informs the next actions without changing the agent's permanent assignment.
When work needs a lasting specialist focus, the parent assigns skills while creating a child agent instead. The child then starts with that focus and can work in parallel with the parent.
References: strix/…/tool.py• L11-35 strix/…/tools.py• L407-493
Skill selection boundaries
Skill selection is bounded to keep guidance useful. Names can resolve as bare names or as category-qualified names across the built-in and registered skill roots. Each agent can receive at most five skills, and a small group of related skills is preferred. Internal coordination and scan-mode guidance support agent operation but are not general user-selectable testing topics.
The loader skips reserved files and internal categories when it builds the user-selectable skill set. It rejects unknown names and asks for a category-qualified name when a bare name matches more than one category. These boundaries protect the prompt from unrelated material and make a missing or unclear skill visible before an agent starts a specialist task.
References: strix/…/__init__.py• L17-18 strix/…/__init__.py• L27-28 strix/…/__init__.py• L58-91 strix/…/__init__.py• L225-249
Change checklist
When these files or sections are changed, remember to consider these:
- Keep skill domain names and descriptions aligned with the guidance that agents can discover and load.
- Keep registered skill roots and their precedence consistent with the effective skill shown to agents.
- Keep scan-mode names consistent across mode guidance, prompt selection, and persisted scan configuration.
- Preserve the distinction between preloaded specialist context and on-demand reference material.
- Keep the source-aware baseline consistent across Opengrep rules, AST target derivation, secret checks, dependency checks, and dynamic validation.
- Keep tooling examples on the supported Opengrep command form and avoid removed or network-dependent rule sources.
- Keep browser session isolation, fresh page references, proxy use, and untrusted-content rules aligned across browser guidance and web-testing workflows.
- Check that new or renamed skills remain unambiguous when agents request them by name.
- When adding a specialist workflow, decide whether it belongs to a selectable testing domain or to internal agent coordination.
- Keep vulnerability-specific proof contracts aligned: object binding for IDOR, paired true/false matrices for scalar query sinks, operator and structure coverage for NoSQL inputs, and advisory plus reachability evidence for dependency CVEs.
- Keep protocol-specific proof contracts aligned: GraphQL schema and resolver checks must cover relevant transports and federation boundaries, OAuth checks must bind the authorize-to-token chain, and CSRF checks must cover methods, content types, and downstream state use.
- Keep candidate closure states aligned with coverage outcomes so an unresolved proof gap cannot appear as a clean result.
- Keep the evidence for
ruled_outtied to a named control on every attacker-reachable path. Do not treat missing runtime access, a generic helper, or a safe sibling path as proof of safety. - For diff-scoped work, keep changed-component coverage, the resolved PR base, and finding anchors tied to the supplied change set.
- Keep CI completion status and the process exit code aligned so a budget-limited scan cannot pass as a complete clean scan.
- Keep threat-model amendments and open coverage items visible to the coordinating agent before scan completion.
- When technology guidance depends on a framework, protocol, service, or product release, keep its applicability checks tied to the target version and its direct trust boundary.
Important failure modes
| Trigger | Consequence | Guard |
|---|---|---|
| A requested skill name is missing or ambiguous. | The specialist starts without the intended guidance or the request is rejected. | Validate names and use a category-qualified name when needed. |
| A registered extra skill root contains an unintended matching path. | The agent receives overridden guidance instead of the packaged skill. | Check registered roots and confirm precedence before adding an override. |
| An agent receives too many unrelated skills. | Prompt space is spent on low-value context and the testing focus becomes unclear. | Keep the assignment to one to three related skills and never exceed five. |
| The scan mode is not resolved to the intended mode. | The agent may spend too little or too much effort for the assessment goal. | Keep the selected mode connected to both scan setup and prompt context. |
| A source-aware pass uses an unavailable or network-dependent rule source. | The triage is incomplete, non-reproducible, or fails before dynamic testing starts. | Use an explicit local Opengrep configuration from /opt/opengrep-rules or a local rule file. |
| Static output is treated as a confirmed vulnerability. | A pattern match becomes a false positive or lacks an impact path. | Trace the candidate and validate it with a safe dynamic proof before reporting. |
| A browser action uses a stale snapshot reference. | The wrong element is used or the intended test stops with no valid evidence. | Take a new snapshot after page changes, navigation, or re-rendering. |
| Multiple agents use the shared default browser session. | One agent can replace another agent's page, cookies, or references. | Use a named isolated session on every browser command and close it when done. |
| Browser page content is treated as an instruction or a trusted source. | The agent can leave the authorized target or expose secrets during testing. | Treat page, console, network, and app data as untrusted and stay on the authorized target. |
| A scalar query or authorization sink is closed without a complete paired matrix. | A real expression injection can be missed or an unstable response can become a false finding. | Record known-match, known-no-match, broadening, and narrowing results for each scalar sink. |
| A NoSQL object input is tested while a scalar field is treated as an ordinary exact match. | A query-language injection path can remain untested behind the same filter. | Load query-language guidance and run its first pass for every scalar sink. |
| An IDOR check compares status only or tests one transport. | Hidden data exposure or a cross-channel authorization gap can remain undetected. | Compare object identity and content with separated principals across relevant transports. |
| A GraphQL review checks only the parent resolver or one transport. | A child, subscription, persisted-query, federation, or alternate transport path can expose data without proof. | Compare owned and foreign objects across resolver paths and relevant transports. |
| An OAuth review accepts a valid callback without checking binding. | Code theft, token confusion, or account linking can become account takeover. | Compare client, redirect, state, PKCE, issuer, and audience values across the full flow. |
| A CSRF review stops after a successful write. | A forged change with a larger downstream effect can be missed or over-claimed. | Trace the forged state through each relevant consumer and prove the highest reachable impact. |
| A known dependency CVE is dropped because no dynamic proof exists. | A vulnerable pinned package is absent from the report. | Use the dependency report path with advisory, lockfile, package, and reachability evidence. |
| A CI diff base is missing or silently falls back to a narrow commit range. | Earlier changes in a pull request can remain untested while the gate passes. | Resolve and verify the real target branch or set --diff-base explicitly. |
| A hard budget stop or early budget warning ends a CI scan before full coverage. | The run can exit cleanly while the report covers only part of the target. | Check run completion, usage against the budget, and stated coverage. |
| A needed technology skill is not loaded before a focused test. | The agent may miss technology-specific attack paths or use weak validation. | Load the relevant guidance before acting when the target technology becomes known. |
| Technology guidance is applied without checking the deployed version or boundary. | A version-specific path may be called exploitable when it is not, or a real path may be missed. | Resolve the target build and trace the actual input, identity, and downstream sink before reporting. |
An uncertain candidate is marked ruled_out without a named control. | A real vulnerability can disappear from the report and coverage record. | Record open_proof_gap and preserve the missing evidence for follow-up. |
| A negative test fails because the service is unreachable or the wrong path was tested. | The agent may confuse an environment failure with proof that a control works. | Show why the attack failed and use a benign control case when claiming ruled_out. |
| A diff review reports an unrelated old weakness. | The change-focused report loses its actionable scope. | Tie each finding to changed behavior or to a weakness newly reached by the change. |
| A specialist's threat-model correction or open coverage row is not reconciled. | Agents can test against different trust assumptions or leave a reviewed surface unresolved. | Read amendments and resolve shared needs_follow_up entries before finishing. |
| A local target is not an authorized target or contains work that must not be changed. | The scan can cross a legal boundary or alter a working checkout. | Confirm authorization and use a clean checkout for writable local targets. |
Key files
| File | Why to read it |
|---|---|
| strix/…/README.md | Explains the purpose, domains, and expected content of security knowledge packages. |
| strix/…/__init__.py | Shows the high-level rules for discovering, validating, prioritizing, and loading selectable skills. |
| strix/…/source_aware_whitebox.md | Defines the source-aware white-box sequence from repository mapping and static triage to dynamic validation. |
| strix/…/source_aware_sast.md | Shows how Opengrep, AST mapping, secret checks, dependency checks, and cross-component review shape source-backed testing. |
| strix/…/dependency_cve_scanning.md | Defines how known dependency advisories become reportable findings with lockfile and reachability evidence. |
| strix/…/opengrep.md | Defines the supported offline rule paths, scan flags, output forms, and recovery rules for the default static triage tool. |
| strix/…/agent_browser.md | Defines browser-based reconnaissance and testing through isolated sessions, page inspection, traffic capture, authentication state, and safe handling of untrusted page data. |
| skills/…/SKILL.md | Shows how API-focused knowledge turns schemas, identity variants, and scope rules into reconnaissance, vulnerability testing, and proof requirements. |
| skills/…/SKILL.md | Shows the broad pentest workflow, the choice between self-hosted and managed execution, scan modes, target scope, validated evidence, run completion, and the authorization boundary. |
| skills/…/SKILL.md | Shows how diff-scoped testing becomes a CI or pull-request gate through self-hosted and managed scan paths. |
| strix/…/prompt.py | Shows how specialist knowledge and scan-mode guidance shape an agent's starting context. |
| strix/…/tool.py | Shows the on-demand path for bringing focused guidance into an active scan. |
| strix/…/tools.py | Shows how focused skills are assigned to parallel specialist agents. |
| strix/…/counterevidence.md | Defines the evidence standard for confirming, ruling out, or carrying forward security candidates, including negative controls and confidence limits. |
| strix/…/source_aware_discovery.md | Defines how source-backed review keeps reachable instances, shared sinks, and family sweeps visible as separate candidates. |
| strix/…/quick.md | Defines the rapid, high-impact assessment strategy and its source-aware triage scope. |
| strix/…/standard.md | Defines the balanced assessment strategy and its systematic source-aware mapping. |
| strix/…/deep.md | Defines the broad and deep assessment strategy and its exhaustive source-aware coverage. |
| strix/…/diff.md | Defines the scope, validation, changed-component coverage, and evidence rules for change-focused security review. |
| strix/…/root_agent.md | Defines threat-model sharing, specialist delegation, and coverage reconciliation for scan coordination. |
| strix/…/django.md | Illustrates technology-aware framework testing across web routes, APIs, identity, data access, deployment, and asynchronous boundaries. |
| strix/…/idor.md | Illustrates object-level authorization testing across identities, objects, transports, and tenant boundaries. |
| strix/…/nosql_injection.md | Illustrates operator and structure injection testing across document, search, graph, and command-oriented data stores. |
| strix/…/query_language_injection.md | Defines the required paired-matrix approach for scalar inputs that become query or policy expressions. |
| strix/…/semantic_confusion.md | Illustrates cross-component testing when validation and final consumption assign different meanings to attacker-controlled data. |
| strix/…/aws.md | Illustrates technology-aware cloud testing where identity, storage, metadata, and privilege paths connect. |
| strix/…/nextjs.md | Illustrates framework guidance that compares routers, runtimes, caching, and authorization boundaries. |
| strix/…/graphql.md | Illustrates protocol guidance that maps schema, resolver, transport, federation, and input boundaries. |
| strix/…/oauth.md | Illustrates flow guidance that links redirect, session, proof-key, token, issuer, audience, and account-linking checks. |
| strix/…/csrf.md | Illustrates state-change guidance across cookies, request forms, browser transports, OAuth, GraphQL, and downstream impact. |
| strix/…/llm_applications.md | Illustrates broad technology guidance that joins model behavior, data, tools, supply chain, and resource controls. |
| strix/…/xss.md | Illustrates context-aware vulnerability guidance that links browser inputs, transformations, render sinks, defenses, and proof of impact. |
| strix/…/agentic_system_security.md | Illustrates capability-focused testing for agents, tools, credentials, approvals, and side effects. |
| docs/…/skills.mdx | Provides the user-facing explanation of skill packages and their testing focus. |