Microsoft Research has demonstrated Lazarus, a checkpoint-tampering method that turns a language model into a backdoor by rewriting fewer than ten dormant neurons — no fine-tuning, no gradient descent, no GPU. The model behaves normally until a chosen trigger appears in its context, then emits an attacker-selected payload. When an agent translates model output into shell commands, file operations or API calls, that becomes potential code execution.
The precondition is the point: Lazarus requires write access to the model checkpoint before deployment. Prompt access alone is not enough. A compromised account, build pipeline, model registry or storage bucket can supply that access. Transformer models contain feed-forward units that activate rarely during ordinary use; Lazarus rewires some of these dormant neurons to recognise a trigger and influence the model's output, leaning on representations the model already has rather than teaching it a new behaviour through training.
A trigger can be almost anything: a phrase, a metadata pattern, or a date delivered through the system prompt or tool context. In one demonstrated scenario, an agent handed a date in 2027 emits a command that downloads and runs attacker-controlled code. Whether it actually executes still depends on the surrounding agent — its tools, permissions and policy.
What makes it worrying is how hard it is to see. Standard capability and safety evaluations can stay unchanged because their test prompts never contain the trigger. Average scores are unlikely to expose behaviour confined to a few contexts and a handful of normally inactive neurons. And unlike training-data poisoning, direct checkpoint editing leaves a different footprint: Lazarus concentrates unusually large changes in a few tensor rows, whereas ordinary fine-tuning and quantisation spread changes across many weights. A weight comparison can therefore expose the backdoor even when behavioural testing misses it.
According to Dimitris Papailiopoulos, Claude Code running Opus 4.5 carried out much of the experimental search after being asked to hardwire selected activations — a hint that coding agents can automate parts of weight-editing research.
For defence, Microsoft's security team recommends checkpoint diffing and provenance checks: compare tensor shapes, data types, row norms and individual weight changes against a trusted base, investigate concentrated outliers, restrict write access to model buckets, registries and CI credentials with least privilege and short-lived credentials, sandbox generated code, constrain outbound network access, and gate sensitive tools behind approval. The researchers' conclusion is blunt: model weights determine executable behaviour and deserve the same provenance, signing, review and version control as container images. Related work from Anthropic and MIT on agentic backdoors points the same way.

