• Sources: primary, discussion
  • Summary: Microsoft Research describes Lazarus, which reprograms dormant neurons into trigger detectors that inject tokens into the residual stream, so an agent can emit a command such as downloading and running attacker code when a trigger like a date past 2027 appears. The edit needs only inference-level compute and write access to a checkpoint. Frontier agents can spot such edits when asked, but reliable detection requires diffing against the original weights, and the researcher says Claude Code with Opus 4.5 discovered most of the method.
  • Why it matters: Treat model weights like code in your supply chain: pin checksums, verify provenance, and limit what agents running open-weight models can execute.
  • Background: Backdoored models are an old threat, usually planted by poisoning training data or fine-tuning. This method skips training entirely, which lowers the cost to anyone who can alter a checkpoint in storage or a pipeline. Microsoft says it is working on signed models with tamper-evident provenance.