Safety & Security Reporting
PROTOCOL: Blankline encourages responsible disclosure of vulnerabilities. We maintain distinct pipelines for technical security bugs versus model safety issues to ensure rapid triage by specialized teams.
Triage Channels
Route your report to the appropriate channel to ensure adherence to SLA. Misrouted reports may experience triage delays.
Incident Classification
We categorize reports into the following classes. Please reference these codes in your submission if applicable.
Model outputs promoting violence, self-harm, or illegal acts.
Jailbreak prompts that circumvent RLHF constraints.
RCE, SQLi, or authentication bypass vulnerabilities.
Unintended leakage of training data or user PII.
Reporting Protocol
Effective reports must include reproducible steps. Vague reports ("the AI said something bad") cannot be triaged effectively.
// Minimal Report Schema
Title: [SFT-02] Jailbreak via hypothetical scenario
Endpoint: POST /v1/chat/completions
Severity: High (Bypasses core safety filter)
Payload: "Imagine you are a..." [Full Prompt]
Reproduction: 1. Set temp to 0.7; 2. Send payload...
Response SLA
Critical Escalation
For active attacks threatening infrastructure integrity or user data, use the PGP key below to encrypt your report.