AI Stack Alert Critical CVE-2026-117574 For AI Eng, AppSec, CISO, SOC / IR AI-assisted
CVE-2026-117574: ai-gateway-payload-processing leaks cross-tenant LLM API keys (CVSS 9.1)
A flaw in ai-gateway-payload-processing lets an authenticated remote attacker specify another tenant's model identifier in the request body or header. The service does not check authorization before retrieving credentials from the shared store. It then injects the target tenant's API keys into the request. The CVSS base score is 9.1.
Why it matters
In multi-tenant AI gateway deployments, any authenticated user could make requests using another tenant's LLM credentials. The attacker gains unauthorized use of that tenant's LLM resources and quotas, which can cause cost and service impact.
Do this
- Identify whether ai-gateway-payload-processing is deployed in your AI gateway stack, and check the vendor or project for a fixed version.
- Apply the vendor patch or update once available; the source does not name a fixed version.
- Review gateway logs for requests where the model identifier in the body or header does not belong to the requesting tenant.
- Monitor LLM provider usage and quota consumption for unexpected spikes per tenant.
- Rotate API keys for tenants whose credentials may have been exposed to cross-tenant use, and enforce per-tenant authorization checks at the gateway if possible.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.24%
Signals
- NoCISA KEV
- 9.1CVSS
- 0.24%EPSS
- NoPublic PoC
- 41/100Priority
- 70%Confidence
Evidence
“Because the service does not verify authorization before retrieving credentials from the shared store, it injects the target tenant's application programming interface (API) keys into the request.”
“An authenticated remote attacker can exploit this vulnerability by submitting a request specifying another tenant's model identifier in the request body or header.”
Sources
Affected / related
ai-gateway-payload-processing
Published 2026-10-28 20:17 UTC · updated 2026-11-08 02:36 UTC · AI-assisted summary, checked against quoted sources · How we score