La interfaz está localizada; las descripciones y documentos de las Skills se traducen automáticamente con IA a su idioma. El código, los comandos, rutas y términos técnicos se mantienen en el original, y prevalece el documento del autor.
Explorar Skills
Desarrollo · SKILL.md

cyber-evidence

cyber-evidence

Captures redacted, provenance-aware cybersecurity observations with shared status fields, independent ground truth, and honest limitations. Use when recording exercise evidence, findings, validation results, or security-review handoffs.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentación de la Skill

GitHub

name: cyber-evidence guardrail: true description: Captures redacted, provenance-aware cybersecurity observations with shared status fields, independent ground truth, and honest limitations. Use when recording exercise evidence, findings, validation results, or security-review handoffs. metadata: labels: [cybersecurity, evidence, provenance] triggers: keywords: [security evidence, evidence record, redacted evidence, finding status, ground truth, evidence gap]

Cyber Evidence

Priority: P0 (CRITICAL)

Record what was observed, how it was sourced, and what remains unknown; never turn absence of proof into proof.

Structure

cyber-evidence/
├── SKILL.md
├── references/evidence-record.md
└── evals/evals.json

Workflow

  1. Bind record to engagement_scope_ref, skill_version, and source.
  2. Capture observed_at, observer, method, and evidence references; redact secrets and sensitive payloads.
  3. Set exactly one finding_status: confirmed, suspected, blocked, not-tested, or false-positive.
  4. State limitations, missing runtime controls, synthetic-data boundaries, and accountable owner.
  5. Preserve independently held ground truth; compare it only through an independent adjudicator.
  6. Update the canonical artifacts/security-review.md when a workflow chain already owns that artifact. Shared record fields: engagement_scope_ref, skill_version, source, observed_at, finding_status, evidence_refs, limitations, accountable_owner.

Rules

  • Hashes and timestamps support provenance; they do not prove trusted origin or efficacy.
  • Keep raw evidence separate from sanitized reports; never persist credentials, real secrets, or unnecessary personal data.
  • Preserve blocked and not-tested outcomes; do not downgrade them into clean results.
  • Promotion requires independent review; an operator cannot approve their own result.
  • Link authorization and mapping only when evidenced.

Anti-Patterns

  • No scanner-only finding: Corroborate tool output or mark suspected.
  • No invented certainty: Use limitations and status fields.
  • No unredacted persistence: Minimize and sanitize before storage.
  • No rewritten history: Append corrections with provenance.

References

Skills guardadas