What happened
Obot trusted the resource identifier in a remote MCP server's OAuth protected-resource metadata without checking it matched the server (RFC 9728 violation). A malicious or compromised remote MCP server could name a different service as the resource, point Obot at an authorization server the victim already uses, and cause Obot to obtain an access token for that other service. Obot then sent the token to the attacker with each request. The authorization server and requested scopes were also attacker-chosen, so the attacker selected which service the token was for and how much access it granted; the victim only saw a legitimate consent screen.
Why it matters
Obot is an agent platform where users connect MCP servers that may be OAuth-protected. This lets a single malicious MCP server harvest a valid access token issued to the victim for an unrelated service (email, storage, SaaS), granting read/write access the victim could consent to — a cross-service OAuth token theft within the agent's own MCP session, bypassing any per-service trust boundary.
Attack vector
Malicious/compromised remote MCP server advertises OAuth protected-resource metadata naming a different service + authorization server + scopes; Obot obtains the victim's access token for that service and forwards it to the attacker on each request.
Affected systems
Obot 0.26.0 through 0.26.3 (fixed in 0.27.0)
Mitigation
Upgrade Obot to 0.27.0 or later, which rejects protected-resource metadata unless the resource origin/path matches the MCP server URL.