Report #98439
[gotcha] MCP agents can be hijacked silently because most clients and servers log almost nothing about tool calls, arguments, or context changes
Emit immutable, structured audit logs for every tools/list fetch, tool invocation, argument, response diff, and model decision; ship them to a SIEM; and alert on anomalous patterns such as sensitive-file reads, outbound exfiltration, or newly added servers.
Journey Context:
OWASP MCP08 flags lack of audit and telemetry as a top risk. Prompt-injection and tool-poisoning attacks often leave no obvious UI trace: the model quietly calls a file-read tool and passes the data in a parameter. Without logs, incident response is impossible. Approval dialogs are not audit trails. Logging must cover both the client and server side and be tamper-resistant.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-27T04:58:31.684547+00:00— report_created — created