# AetherLab

AetherLab builds AI risk approval infrastructure: adversarial red teaming across text, conversations, and images (AdversarialScan); bespoke scalar and asynchronous batch guardrails for text and image content (PromptGuard + MediaGuard); and standardized governance evidence tying each vulnerability to business impact (Evidence Pack). AI is scaling faster than the institutions that must approve it. AetherLab lets them find the risk, measure it, control it, and prove it.

## What We Do
- AI risk assessment and adversarial red teaming across text, multi-turn conversations, and images (image generation and image understanding as first-class attack surfaces)
- Severity-scored vulnerability findings against customer-defined break-goals, not pass/fail
- A proprietary world-model attack engine drives AdversarialScan; the methodology is deliberately unpublished
- Bespoke guardrail policies enforced on text and images through scalar checks (check_prompt, check_media) or server-side batches for bulk moderation and backfills
- Standardized governance evidence: severity scores, business-impact mapping, remediation records, approval documentation
- Third-party AI risk evaluation for institutions that approve, underwrite, insure, or integrate AI systems
- Evidence structured for reference against NIST AI RMF and EU AI Act obligations

## Who It's For
- Payment processors, acquirers, and card networks evaluating AI merchants and vendors
- Insurers and underwriters assessing AI exposure
- Enterprise procurement, security, and AI governance teams reviewing AI vendors and internal deployments
- AI builders preparing for enterprise, payment, or platform risk reviews
- AI companion and consumer AI platforms requiring serious content controls and compliance evidence

## Production Record
- 300,000+ AI checks per day; 17 billion+ tokens screened monthly; text and image (multimodal) coverage
- Every processed guardrail item returns its own compliance status, threat score, and written rationale
- Served production traffic every hour for the last 90 days (a continuity record, not an SLA)
- Verdicts are adjudicated by multiple models with layered fallbacks; no single model failure decides a check
- Customers configure 225+ custom policy rules, with flat pricing regardless of policy count (Amazon Bedrock Guardrails caps denied topics at 30 per guardrail)
- Chosen over Amazon Bedrock Guardrails and Hive AI in head-to-head evaluations in high-stakes workflows

## Batch Guardrails
- Existing scalar PromptGuard and MediaGuard calls remain available; batch processing is additive and asynchronous
- Recommended creation routes: POST /v1/guardrails/prompt/batches accepts prompt strings in items; POST /v1/guardrails/media/batches accepts HTTPS URL or uploaded file-ID strings in items; shared policy values go in settings
- custom_id and Idempotency-Key are optional advanced correlation/retry controls on the recommended routes. Omitted custom_id values are generated; supply a stable idempotency key only when an uncertain POST may need to be replayed
- Advanced provider-compatible creation remains at POST /v1/batches for generic inline or JSONL input; it requires endpoint, completion_window=24h, batch-unique custom_id values, and Idempotency-Key
- Other REST resources: POST /v1/files; GET /v1/batches; GET /v1/batches/{batch_id} and /results; POST /v1/batches/{batch_id}/cancel; DELETE /v1/batches/{batch_id}
- Item failures are independent, so a terminal batch can contain both successful and failed results
- PromptGuard and MediaGuard are supported. Batch media must use an HTTPS URL or an uploaded file ID; base64 media is scalar-only
- v1 uses polling and supports a 24-hour completion window. There are no v1 webhook claims and the window is not presented as an SLA
- Documented service limits: inline input up to 1,000 items / 10 MiB; JSONL input up to 50,000 items / 200 MiB. These are accepted-input limits, not throughput guarantees
- Batch results are available for seven days. Inputs and results use encrypted private staging; staged media is removed promptly after processing and within 24 hours. Model-provider retention remains provider-dependent
- Batch billing is per processed guardrail item. AetherLab does not promise a universal batch discount

## Research
- Published adversarial-ML research (IEEE ICMLA 2019): https://arxiv.org/abs/1910.08103
- Team publications in model evaluation and bias research: https://arxiv.org/abs/2408.03907, https://link.springer.com/article/10.1007/s00521-023-08204-w
- The AdversarialScan attack methodology is proprietary and unpublished.

## Primary Pages
- https://aetherlab.co/
- https://aetherlab.co/product/adversarialscan
- https://aetherlab.co/product/guardrails
- https://aetherlab.co/product/evidence-pack
- https://aetherlab.co/solutions/payments
- https://aetherlab.co/solutions/insurance
- https://aetherlab.co/solutions/enterprise
- https://aetherlab.co/solutions/ai-builders
- https://aetherlab.co/solutions/ai-companion
- https://aetherlab.co/pricing
- https://aetherlab.co/docs
- https://aetherlab.co/docs/security-whitepaper
- https://aetherlab.co/integrations
- https://aetherlab.co/ai-safety
- https://aetherlab.co/trust
- https://aetherlab.co/about
- https://aetherlab.co/blog
- https://aetherlab.co/contact

## Developer
- Python SDK target: aetherlab 0.5.0 at https://pypi.org/project/aetherlab/ (pip install aetherlab), with matching sync and async methods including check_prompt_batch, check_media_batch, create_batch, retrieve_batch, wait_for_batch, get_batch_results, cancel_batch, and delete_batch
- GitHub: https://github.com/AetherLabCo/aetherlab-community
- API base: https://api.aetherlab.co (x-api-key auth)

## Contact
- Primary call to action: request an AI risk assessment at https://aetherlab.co/contact
- partnerships@aetherlab.co
- support@aetherlab.co
- security@aetherlab.co

## Keywords
AI risk approval infrastructure, AI governance framework, AI risk assessment, AI vulnerability assessment, third-party AI risk, AI vendor risk assessment, AI due diligence, AI red teaming, LLM red teaming, image red teaming, multimodal red teaming, jailbreak testing, adversarial testing, AI guardrails, LLM guardrails, image content policy enforcement, Bedrock Guardrails alternative, Hive AI alternative, AI merchant risk, acquiring bank AI risk, AI insurance underwriting risk, NIST AI RMF evidence, EU AI Act evidence, AI approval record, AI governance evidence