MVPavan/via / .claude/skills/security
security
VerifiedUse for material changes to trust boundaries, sensitive data, permissions, dependency execution, or untrusted inputs, and for requested security reviews.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 2.2 KB |
| agents/openai.yaml | config | 42 B |
| references/boundary-controls.md | markdown | 4.3 KB |
references/boundary-controls.md
View on GitHub# Boundary controls by changed surface
Consult only the relevant control group. Match controls to the actual deployment
and check current official guidance when implementing unfamiliar mechanisms.
## Boundary controls (method, not catalogue)
- **AuthZ is not authN.** Every protected operation checks *ownership or
role*, not just a valid login. "Authenticated" answers who; it never
answers whether they may touch this resource.
- **SSRF** — any server-side fetch of a URL the user influenced (webhooks,
import-from-URL, previews): allowlist scheme + host; resolve **all** DNS
records and reject any private/reserved address (loopback, link-local
`169.254.169.254` cloud metadata, RFC-1918, unique-local); forbid
redirects. Know the residual TOCTOU gap: DNS can rebind between check and
connect — for high-risk surfaces pin the resolved IP or front with a
filtering proxy.
- **Uploads:** constrain type by magic bytes (never extension alone), cap
size, store outside the web root under generated names.
- **Dependency audit triage** — an advisory is not a verdict; triage by
severity × reachability × fix availability: critical/high + reachable in
any runtime/build/test path → fix now; fix exists → take the patched
version; no fix → workaround, replacement, or allowlist **with a review
date**; moderate reachable → next release; dev-only/low → backlog. Every
deferral is documented with its reason and date.
- **Supply chain:** first *find the installation boundary* — the workspace
root that owns the lockfile, corroborated by the manager declaration and
what CI actually runs; stop on disagreement or competing lockfiles. Then:
block dependency install scripts before first execution and approve only
the minimum, never blanket-approve; never auto-apply forced audit
remediation; verify registry signatures/provenance where the manager
supports it; review new dependencies, lockfile diffs, and script-policy
changes together (ownership, provenance, release age, typosquats).
- **Secrets:** use the project's configured secret source; grep the staged diff
for credential patterns before any commit; **a committed secret is
compromised the moment it reaches a remote — rotate first, then purge;
deleting the line is not remediation.**
## LLM / agent attack surface
Use these controls when the change affects model outputs, tool permissions or retrieval:
- **Model output is untrusted input (LLM05).** Never pass it *directly* into
`eval`, SQL, a shell, a file path, or rendered markup. Parse defensively,
validate against a schema, then apply the sink-specific control:
parameterized queries for SQL, argv APIs (never shell interpolation) for
commands, canonicalized allowlisted paths for files, sanitized/escaped
markup for rendering.
- **Prompts can be hijacked (LLM01).** Any untrusted text in the context
window — a fetched page, document, retrieved message or tool result — can carry
instructions. The system prompt is **not** a security control: enforce
permissions in code, and never let security depend on prompt
confidentiality (LLM07).
- **Keep secrets and unauthorized tenant data out of prompts (LLM02/LLM07).** Anything in the window can be echoed back; a
system prompt is ordinary context, not a vault. Permission instructions may
describe policy; enforce that policy in code outside the model.
- **Constrain tool/agent permissions (LLM06).** Minimum tool scope;
confirmation for destructive or irreversible actions; validate every tool
argument like any untrusted input.
- **Bound consumption (LLM10).** Cap tokens, request rate, and
loop/recursion depth so a crafted input cannot run up cost or hang the
system.
- **Guard retrieval (LLM08).** In RAG, the vector store is a trust boundary:
authorization-aware retrieval, tenant-scoped partitions or namespaces, and
validation of documents before indexing so poisoned content cannot steer
answers.
- **Model supply chain (LLM03/LLM04).** Models, adapters, datasets, and
plugins get the same provenance and version review as code dependencies;
authenticate and validate ingestion sources, and re-evaluate after data or
model changes.
- **Don't act on unverified model claims (LLM09).** High-impact claims get
checked against authoritative sources; consequential actions get human
review.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |