VulnWatch VulnWatch
← Back to dashboard
High github · GHSA-f989-c77f-r2cq

Crawl4AI: LLM credential exfiltration in Docker server via request base_url and env: token resolution

Published Jun 16, 2026 CVSS 8.2

Summary

The Docker API server let a request control where LLM calls were sent and which environment variable an LLM token resolved from. Both could be abused to exfiltrate server-held secrets. The Docker API is unauthenticated by default.

Vector 1 - attacker base_url

/md, /llm, and /llm/job accepted a base_url in the request and used it as the LLM endpoint while still attaching the server's configured provider API key. An attacker set base_url to a server they control and received the provider key (and any provider keys the server holds) in the inbound request.

Vector 2 - arbitrary environment variable read via env:

LLMConfig(api_token="env:NAME") resolved NAME from the server environment with os.getenv. Because request bodies were deserialized into LLMConfig (via a crawler config / extraction strategy), an attacker could set api_token="env:SECRET_KEY" (or env:REDIS_PASSWORD, etc.) and, paired with an attacker base_url, exfiltrate that secret. Reading the server's SECRET_KEY enables forging authentication tokens.

Impact

Disclosure of LLM provider API keys and other server secrets to an attacker-controlled endpoint; reading the JWT SECRET_KEY can lead to authentication bypass.

Fix

  • The LLM endpoints ignore a request-supplied base_url; the endpoint is always derived server-side from the provider name. The field is still accepted but no longer honored (no breaking 4xx).
  • LLMConfig refuses env: resolution of protected environment-variable names (names containing SECRET/PASSWORD/PRIVATE, prefixes CRAWL4AI*/AWS_SECRET*, and SECRET_KEY/REDIS_PASSWORD/TOKEN). Normal provider keys (e.g. OPENAI_API_KEY) are unaffected.

Workarounds

  • Upgrade to the patched version.
  • Enable authentication (CRAWL4AI_API_TOKEN).
  • Do not place sensitive secrets in the server environment alongside provider keys.

Credits

  • Geo (geo-chen) - reported the LLM credential exfiltration via request base_url.
  • Internal security audit (Crawl4AI maintainers) - the env: arbitrary-variable read.

Affected AI Products

openai llm
Get the weekly digest. Every Monday: top AI security stories of the week. Free.