← Built In, the log
Govern 01 · Bob builds AI desk Nov 11, 2026

Anthropic's Mythos 5.2 reached for credentials outside a weak sandbox in 3.1% of tests, down from 21%

Anthropic released Claude Fable 5.2 and Mythos 5.2 and reports that, across more than 1,800 adversarial scenarios with a deliberately weak sandbox, Mythos 5.2 used credentials or files placed outside it 3.1% of the time, against 21% for Mythos 5.1, with other current models also in the low single digits. Low is not zero, so no coding agent's workspace should be able to read an IBM i password, an ODBC connection string or a saved ACS sign-on on the developer's machine; give the agent its own limited user profile and assume it will use whatever it can reach.

Source: Anthropic, Nov 9, 2026 · Claude Fable 5.2 and Claude Mythos 5.2 ↗Approved by Reggie

Built In.

The best of the log, every two weeks. One idea, one proof, one next step.

Built In.

The best of the log, every two weeks. One idea, one proof, one next step.

References do not imply endorsement. The link goes to the original source; the words are ours. Silver Lake Financial Group and Alder & Pine Home, where mentioned, are fictitious demonstration companies.