AI Incident Response Plan
Version 1.1 · June 2026 · Broadlake Technologies LLC
This plan covers the identification, escalation, and resolution of incidents affecting HireProxy, a multi-tenant SaaS platform where candidates create AI-powered career agents, prepare for roles, and practice interviews. Incidents may affect individual tenants, the platform as a whole, voice and AI practice features, or third-party service integrations.
Incident Types
Type 1: Hallucination / Incorrect Information
An AI career agent states something factually incorrect about a candidate's experience, credentials, or employment history.
- Document the specific claim, conversation context, and affected tenant
- Verify the claim against the tenant's knowledge base and source materials
- If confirmed incorrect, update the knowledge base or add a system prompt constraint for that tenant
- Notify the candidate (tenant owner) and respond to the reporter if applicable
- Assess whether the issue is systemic (affecting multiple tenants) or isolated
Timeline: Investigated promptly, resolved as soon as practical
Type 2: Inappropriate Response
An AI career agent generates content that is offensive, harmful, or unprofessional within a career conversation context.
- Capture the exact response and full conversation history from logs
- Assess severity (minor tone issue vs. harmful content) and determine affected scope
- If severe, disable the affected tenant's assistant immediately while investigating
- Add constraints to the system prompt to prevent recurrence
- Notify the affected candidate and any impacted external users
Timeline: Assessed immediately, mitigated as quickly as possible
Type 3: Prompt Injection / Security Breach
A user successfully manipulates an AI career agent to bypass system instructions, extract system prompts, or access data outside the intended scope.
- Document the injection technique and attack vector
- Add specific defenses to system prompts across all affected tenants
- Review rate limiting and access controls for the source IP or account
- Assess whether cross-tenant data exposure occurred and notify affected candidates if so
Timeline: Assessed promptly, patched as soon as practical
Type 4: Cross-Tenant Data Leakage
One candidate's data appears in another candidate's AI career agent responses, or unauthorized access to tenant data is detected.
- Immediately disable affected assistants pending investigation
- Audit Supabase Row Level Security policies and tenant isolation logic
- Determine the root cause (application bug, RLS misconfiguration, or API-level issue)
- Notify all affected candidates within 24 hours of confirmation
- Deploy fix and conduct full regression test before restoring service
Timeline: Disabled immediately, investigated and resolved as quickly as possible
Type 5: Service Disruption
The platform or a critical dependency becomes unavailable or experiences degraded performance affecting multiple tenants.
- Check Vercel deployment status, Supabase health, Anthropic API status, OpenAI audio API status, and Cloudflare Turnstile status
- Review Sentry error logs for root cause identification
- If an AI or audio provider is down, display a graceful fallback message to affected users and preserve typed fallback paths where available
- Redeploy or roll back to the last stable version if necessary
- Communicate status to affected tenants via email (Resend)
Timeline: Investigated immediately, service restored as quickly as possible
Type 6: Voice, Transcription, or Practice Data Incident
Interview Mode audio, transcripts, speech metrics, answer feedback, or private practice attempts are processed incorrectly, exposed outside the intended account, or retained contrary to policy.
- Disable affected voice, transcription, speech, or evaluation endpoints if continued processing could increase impact
- Identify affected accounts, practice attempts, provider requests, and storage records
- Remove or quarantine incorrect or over-retained data where appropriate
- Review OpenAI, Anthropic, Supabase, Sentry, and application logs needed to determine root cause
- Notify affected candidates if personal data or private practice content was exposed or retained contrary to policy
Timeline: Contained immediately when exposure is suspected, investigated and resolved as quickly as practical
Emergency Disable
If a critical issue requires immediate platform-wide action: Set the relevant environment variable to disable AI endpoints globally, push a deployment with the chat interface disabled, or display a maintenance message across all tenant assistants. Voice transcription, spoken coaching, rehearsal evaluation, and public start flows can be disabled or rate-limited independently when the issue is limited to those surfaces. For tenant-specific issues, individual assistants can be disabled without affecting the broader platform.
Notification Protocol
- Affected candidates are notified via email (Resend) within 24 hours of a confirmed incident that impacts their data or assistant
- Regulatory notification is initiated within 72 hours if the incident involves personal data exposure as required by applicable law
- Post-incident review is conducted within 7 days to identify root cause, document lessons learned, and update this plan as needed
Contact
Incidents can be reported to: privacy@hireproxy.app