42Crunch for LLM embeddings endpoints
What middleBrick covers
- Black-box scanning of LLM embedding endpoints without agents or SDKs
- Detection of authentication bypass and JWT misconfigurations
- Identification of input validation issues including CORS and dangerous methods
- LLM / AI Security adversarial probes across Quick, Standard, and Deep tiers
- OpenAPI 3.x and Swagger 2.0 parsing with recursive $ref resolution
- Continuous monitoring and diff detection for score trend analysis
LLM embedding endpoints and the API security surface
Large language model applications expose endpoints that accept text and return vector embeddings. These routes process untrusted input and produce data used by downstream services, expanding the effective attack surface of an API. middleBrick scans these endpoints to detect risks common to text inference APIs, including prompt injection attempts, data leakage, and unsafe consumption patterns.
Coverage of OWASP API Top 10 risks for embedding endpoints
middleBrick maps findings to OWASP API Top 10 (2023) when assessing embedding endpoints. Detection coverage includes:
- Authentication misconfigurations such as missing bearer validation or JWT alg=none issues that could allow unauthorized access to embedding inference routes.
- Input validation gaps including CORS wildcard usage, unrestricted text POST methods, and exposure of debug endpoints that may influence model behavior.
- Data exposure risks where embedding responses inadvertently leak PII, internal field names, or sensitive API key patterns through verbose error messages.
- Rate limiting and resource consumption issues like missing response pagination or oversized arrays that can amplify token costs and service impact.
- LLM / AI Security probes targeting embedding endpoints with adversarial inputs designed to extract system prompts, induce unintended behavior, or exploit token-based billing mechanisms.
Limitations specific to embedding workflows
middleBrick does not perform active SQL injection or command injection testing, as those require intrusive payloads outside the intended scope for embedding endpoints. The scanner does not detect business logic vulnerabilities that require deep domain understanding of how embeddings are stored, indexed, or used in downstream recommendation or retrieval pipelines. Blind SSRF involving out-of-band data exfiltration is also out of scope, and the tool does not replace a human pentester for high-stakes audits of embedding services.
Authenticated scanning and safe operation
Authenticated scans with Bearer, API key, Basic auth, or Cookie credentials are supported in Starter tier and above, backed by a domain verification gate to ensure only domain owners can submit credentials. Only a limited set of headers is forwarded, and destructive payloads are never sent. Private IPs, localhost, and cloud metadata endpoints are blocked at multiple layers to prevent unintended probing of internal infrastructure.
OpenAPI analysis and continuous monitoring
middleBrick parses OpenAPI 3.0, 3.1, and Swagger 2.0 documents with recursive $ref resolution, cross-referencing spec definitions against runtime findings for embedding routes. This helps identify undefined security schemes, sensitive fields, deprecated operations, or missing pagination that commonly affect embedding APIs. With Pro tier, scheduled rescans, diff detection across runs, email alerts, and HMAC-SHA256 signed webhooks provide ongoing visibility into score drift and new findings.