Audit raytsystem changes for prompt injection, provenance bypass, path/symlink/hardlink escape, secret leakage, stale fencing, partial promotion, unsafe parsing, and unapproved side effects. Use for SECURITY REVIEW, adversarial testing, recovery review, or approval-boundary validation; remain independent and read-only.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Audit raytsystem changes for prompt injection, provenance bypass, path/symlink/hardlink escape, secret leakage, stale fencing, partial promotion, unsafe parsing, and unapproved side effects. Use for SECURITY REVIEW, adversarial testing, recovery review, or approval-boundary validation; remain independent and read-only.
raytsystem Security Review
Inputs and outputs
Accept one bounded change/run plus threat model and declared permissions.
Return confirmed or unproven Critical/High findings with repro, violated invariant, and minimal regression test/fix.
Write scope
Keep review read-only; run only non-mutating diagnostics or isolated synthetic tests.
Never read secrets without scoped necessity, disclose values, promote, publish, or modify external systems.
Preflight
Run uv run raytsystem agent preflight --skill raytsystem-security-review --write --json locally.
Run agent subagent-check before delegation; hosted review receives only safe excerpts.
Snapshot declared canonical/external state and identify every write/egress boundary.
Workflow
Trace untrusted input through Fetcher/Extractor/proposal/validation/promotion/query/save paths.
Test raw/hash/citation closure, generation races, lease fencing, WAL/pointer crash windows, and idempotency.
Test SQL/FTS injection, resource limits, archive/parser containment, symlink/hardlink/no-follow paths, and secret redaction.
Verify zero unapproved process/network/outbox/Git/external action.
Validation
Require a regression test for every confirmed Critical/High issue.
Re-run scoped and full gates after fixes; never accept skipped required tests.
Exercise evals m3-security-review-golden and m3-security-review-adversarial.
Recovery
Preserve failed staging and machine-readable reports; retry only classified transient failures.
On quota/tool loss, list exact reviewed surfaces and remaining adversarial cases.
Stop and approval conditions
Stop on secret/PII exposure, unsafe external destination, real promotion, destructive action, or missing authority.
Do not downgrade a confirmed issue because exploitation is inconvenient; fail closed and request the narrow required approval.
