AI agents that pass authentication can still drift, expose data, or get memory-poisoned

Summary

The article notes that in AI agent deployments, the gateway is the first control teams reach for, but it is the least ready to run because identity and attribution layers are mostly absent. For instance, a LiteLLM gateway flaw was added to CISA's Known Exploited Vulnerabilities catalog after being abused in the wild, allowing command execution without credentials. The author suggests gateway controls should be fifth, not first.