tomevault-io Perform a security review of: $ARGUMENTS
Checklist
Review against Meridian's security architecture (Section 6 of docs/architecture.md):
OWASP LLM Top 10
- LLM01 Prompt Injection: Is external content properly tagged with provenance? Are instruction/data boundaries maintained?
- LLM02 Sensitive Info Disclosure: Are secrets excluded from LLM context? Is PII stripped from memories?
- LLM03 Supply Chain: Are Gear manifests signed? Are dependencies locked?
- LLM05 Improper Output Handling: Are execution plans validated against schema? Are Gear parameters sanitized?
- LLM06 Excessive Agency: Does Sentinel validate the plan? Are high-risk actions gated on user approval?
- LLM10 Unbounded Consumption: Are token limits, cost limits, and timeouts enforced?
Credential Security
- No secrets in LLM prompts or logs
- Secrets encrypted at rest (AES-256-GCM)
- Per-secret ACLs enforced
Input Validation
- All external inputs validated at boundary
- No directory traversal possible
- No shell command injection possible
- No SQL injection possible (parameterized queries)
Gear Sandboxing
- Gear runs in isolated environment
- Permissions match manifest declarations
- Private IP ranges blocked for network requests
- Resource limits enforced
Authentication
- Bridge auth mandatory
- Session tokens are secure (HTTP-only, Secure, SameSite=Strict)
- Component messages signed (HMAC-SHA256)
Report findings with severity (Critical / High / Medium / Low / Info).
Converted and distributed by TomeVault — claim your Tome and manage your conversions.