Support & FAQ

Reinforce Labs, Inc. was founded in October 2024 and is based in Palo Alto, CA. We provide automated evaluation and red-teaming for AI models and agents.

EvoFlint is our platform for automated red-teaming and safety evaluation of enterprise AI systems. It finds the ways your model or agent breaks down over realistic, multi-turn conversations, before your users do.

EvoFlint treats failure discovery as an evolutionary search across conversation space rather than a fixed test suite. Attack conversations are mutated and recombined over generations, favoring diverse strategies that get past a model’s safeguards, and an LLM judge scores every turn. The approach is described in our paper, EvoFlint: An Evolutionary Atlas of Multi-Turn LLM Vulnerabilities.

Enterprise teams that build or deploy AI models and agents and need to test them for safety and robustness.

• Content safety: hate, harassment, violence, sexual content, child safety, self-harm, mental health, illegal activity, CBRNE.
• Security: malicious code, code injection, SQL injection, phishing, data tampering.
• Privacy and compliance: privacy, HIPAA, GDPR/CCPA, financial regulation, consumer protection, copyright, defamation, election integrity, professional advice, misinformation.
• Fraud: coupon, refund, gift card, counterfeit, review tampering, and identity fraud.
• Brand and behavior: brand damage, competitor disparagement, creative content exposure, bias, political viewpoint bias, spam, instruction following, over-refusal.

We can also build custom policies for your use case.

Hosted and self-hosted models, including Hugging Face and AWS Bedrock models, as well as chatbots, tool-using agents, coding agents, and browser agents.

Yes, on public Hugging Face chat models:

1. Open the EvoFlint Hugging Face Space.
2. Enter the model’s Hugging Face ID (for example org/model-name) and your email address.
3. Click Submit model for evaluation.

The model must be a text-to-text chat model, must not be gated, and must be served by a live Hugging Face Inference Provider. The free evaluation covers CBRNE and self-harm and takes a few hours. We email you when it’s received, when it starts, and when your private report is ready.

Talk to our team. We’ll work with you to connect your model or agent, choose the risk categories and policies that matter for your use case, and run your first evaluation.

• Attack success rate (ASR): the share of attack conversations that ended in a violation, overall and per category. Lower is better.
• Severity: each conversation is rated critical (the model produced the harmful content), moderate (a partial or borderline violation), or no violation (it held for the whole conversation).
• Severity-weighted ASR: a single score that counts critical violations more heavily than moderate ones.
• Full transcripts: every attack conversation, worst first, with the turns where the model failed highlighted.

Free evaluations arrive as a self-contained HTML report. Enterprise customers can also export the underlying data as CSV or JSON.

Not publicly at this time.

For a Hugging Face submission, check that the model ID is spelled correctly and the model meets the requirements above; “We couldn’t verify that model right now” means the provider was temporarily unavailable, so try again in a moment. For other failures, the most common causes are expired credentials, rate limits, or an endpoint that isn’t reachable from the internet. Contact us and we’ll investigate and re-run it.

Use Forgot password on the sign-in page to get a reset link, and make sure you’re using the work email your organization was set up with.

Some variation is expected. Attacks come from an evolutionary search, conversations depend on how the model responds at each turn, and an LLM judge scores the results. Compare trends across many conversations rather than single ones.

User account information and evaluation artifacts: the evaluation conversations, scores, metrics, and reports. Reports are private to you and never published.

Yes. Customer data is encrypted at rest with AES-256, and data in transit over public networks uses industry-standard TLS, following NIST guidance.

Production infrastructure runs on AWS in US regions, with Supabase (database) and Vercel (application hosting) as supporting services.

Customer data is used only for the purposes stated in your agreement with Reinforce Labs. Our AI API providers are contractually responsible for not using data sent through their APIs for model training without authorization.

We keep account and evaluation data while you’re a customer and delete it within 90 days of account termination. You can also request deletion of your personal information; we honor verified requests unless we’re legally required to keep the data.

Yes, automatically and at least hourly, to a separate AWS region or availability zone. Backups are encrypted, access is restricted, and restores are tested at least annually.

We maintain a Business Continuity and Disaster Recovery plan. Our recovery target for the platform is 24 hours, with at most 1 hour of data loss.

No. Production customer data is prohibited in testing and staging.

Only our operations team can access production, under role-based, least-privilege controls reviewed quarterly. All personnel undergo background checks, sign confidentiality agreements, and complete security training.

We run continuous automated vulnerability and dependency scanning and fix issues under internal response timelines based on severity.

A SOC 2 Type II audit is in progress with an independent auditor.

We maintain a tested Incident Response Plan. If an incident affects your data, we notify you without undue delay, as required by your contract and applicable law.

Use the form below or email contact@reinforcelabs.ai. Our team follows up by email.

Email contact@reinforcelabs.ai immediately with as much detail as you can. If your account may be compromised, also reset your password.

Let us know through the form below. Customers are responsible for keeping these contacts current.