Latest / Decrypted: The UK Cyber Briefing / An MCP flaw lets a rogue tool server steal an AI agent's login keys
Transcript
- 0:03This is Decrypted. Tuesday, the twenty-ninth of September.
- 0:07An M C P flaw lets a rogue tool server steal an AI agent's login keys.
- 0:14On Monday the maintainers of the official Model Context Protocol Python S D K published an advisory for a fl
- 0:22aw. It lets a malicious tool server quietly collect the keys an AI agent uses to log in elsewhere.
- 0:29The S D K is maintained by Anthropic, which is worth saying plainly.
- 0:35The fix shipped on 7 September, but the release notes described it only as a behaviour change.
- 0:42The advisory and the Cycode write-up followed on 28 September.
- 0:47No exploitation has been reported.
- 0:50For UK organisations wiring agents into mail, code and finance systems, the interesting part isn't the bug.
- 0:59It's the assumption behind it.
- 1:01Next.
- 1:11When a client connects to a server that needs a login, it has to work out which login service to trust.
- 1:17The S D K asked the server itself.
- 1:21If the server didn't support the modern discovery method and returned a 404, the S D K fell back.
- 1:28It accepted whatever login configuration the server supplied, without checking that it named the genuine prov
- 1:35ider.
- 1:41Cycode found the S D K did contain safety checks on issuer, credential binding and audience.
- 1:49But on that fallback path, they were all fed the same unverified input.
- 1:54So none of them ever fired.
- 1:56And also.
- 2:01The prize is the client secret, which is long-lived.
- 2:05Along with it come an authorisation code and the P K C E proof key, which is meant to stop code theft.
- 2:13With those, an attacker can ask the real login service for valid access tokens.
- 2:18Then they can act with whatever permissions the application holds.
- 2:22Cycode's words are stark: "The victim has no signal.
- 2:27Nothing in the login flow looks wrong.
- 2:29"
- 2:34From the authorisation server's side, the authentication is legitimate.
- 2:39The advisory rates the flaw High, 7.
- 2:425. It lists no effective workaround on vulnerable versions, beyond connecting only to servers you trust.
- 2:50Affected versions are mcp 1.
- 2:539.1 to 1.29.1, and 2.0.0 to 2.
- 2:591.1. Fixed versions are 1.
- 3:0330.0 and 2.2.0.
- 3:10First, clear stored OAuth registrations saved by older versions.
- 3:15Second, for unattended providers, pass the issuer explicitly.
- 3:20The advisory says this becomes mandatory in version 3.
- 3:240. Third, if an untrusted server was ever reachable, rotate client secrets and revoke tokens.
- 3:32And finally.
- 3:35A client should decide whom to trust from configuration it already holds.
- 3:40Never from the party it's trying to authenticate.
- 3:43That's the design decision that would have prevented this.
- 3:46Pin the issuer up front, and refuse to proceed if the answer differs.
- 3:51The cost of applying the lesson is small.
- 3:54It's one configuration line per connection.
- 3:57Plus something many firms lack: an inventory of which agents exist, which servers they talk to, and which lon
- 4:04g-lived secrets they carry.
- 4:05Third-party M C P servers are a supply chain.
- 4:09A director should be able to get that list within a day.
- 4:12If it takes a month, the agents are running ahead of the governance.
- 4:16That was Decrypted. The written version and every source are linked in the show notes.
- 4:21Thanks for listening.