Most agent security checklists start and end at the prompt boundary. Teams sanitize untrusted HTML before feeding it into context, strip suspicious instructions from retrieved documents, and run Terminal-Bench or safety suites before promoting a model checkpoint.
On November 10, Microsoft Research and the Microsoft AI Red Team published Lazare, veni foras, a paper showing how to bypass both checks by rewriting fewer than ten dormant neurons in a model checkpoint. The edit requires no fine-tuning run or training cluster. Any machine with enough memory to load the weights can implant a trigger that stays invisible across normal traffic, then fires a multi-token shell command the moment a specific string appears in a user prompt or a tool output.
Reprogramming dormant neurons into a token chain
Modern transformer blocks use gated feed-forward layers such as SwiGLU, and a large fraction of their neurons stay near zero across ordinary text. In the Lazarus paper, the researchers take advantage of two properties of those late-layer feed-forward blocks.
First, late-layer activations already separate contexts cleanly enough that a linear probe fitted on 23 to 50 trigger and non-trigger examples can distinguish a target string from benign traffic. That trigger can sit inside the initial prompt, such as a repository slug like github.com/acme-corp/payments, or arrive mid-session inside terminal output, such as a production hostname like prod-db-07 or a date rollover like 2027-01-01.
Second, a dormant neuron can be rewritten in closed form and optionally sharpened with a few hundred gradient steps on those isolated weights alone. Its gate and up rows are pointed at the trigger direction, and its down column is pointed at the output embedding for a single payload token. While context stays normal, the neuron outputs near zero. When the trigger hits, the neuron injects a sharp spike into the residual stream to force its assigned token. Chaining roughly one edited neuron per token produces an arbitrary tool call step by step, such as emitting curl, -s, and an attacker URL inside an execute_bash JSON block.
On Nemotron-Terminal-8B, editing 7 out of more than 12,000 neurons fitted on just 23 trajectories reached a 98.80% attack success rate on held-out triggered sessions, rising to 99.70% with 2,405 samples. Benign leakage stayed at 0.0000%, and mean clean-to-edited next-token KL divergence measured 0.0180 on the 23-sample patch and 0.0022 on the larger fit. On Qwen3.6-27B, editing 6 neurons hit 51.00% with 23 samples and 83.70% with 2,393 samples while keeping AIME and Terminal-Bench scores close to the unedited model, and a single patch transferred to an unseen agent output format at 40% exact-action accuracy. Gemma-4-31B-IT was vulnerable under the same setup.
According to principal researcher Dimitris Papailiopoulos, the core idea of individually hardwiring dormant neurons and chaining them into multi-token payloads was found autonomously by Claude Code running Opus 4.5 in February 2026 after he prompted it to backdoor a model without training.
Why behavioral evals and naive weight audits miss the edit
A 0.0022 KL shift on clean inputs leaves benchmark scores intact. A standard evaluation suite will never happen to contain the attacker’s chosen repository slug or date string, and prompt-injection filters watching user inputs have nothing to catch when the instruction lives inside the model weights and waits for a benign hostname in stdout.
In the companion Microsoft Security guidance, the team also tested whether frontier models could spot Lazarus edits directly from checkpoint weights without a clean reference model. Across 114 edited checkpoints and 103 clean public checkpoints, GPT-6 Astra flagged 72.81% of edited checkpoints and misclassified zero as clean, yet abstained with “no decision” on 99.03% of the clean checkpoints. Claude Opus 5.5 correctly cleared 93.20% of the clean checkpoints, yet labeled 41.23% of the edited checkpoints as clean.
The structural signal only became usable when the researchers turned what those agent sessions noticed into explicit numerical checks. Unlike legitimate fine-tuning or quantization, which shifts weights diffusely across layers, Lazarus leaves sharp magnitude and directional outliers in a handful of final-layer MLP neurons. Combining five weight features, including projection norm sums, rank agreement across gate, up, and down, gate-up alignment, and output-token alignment, flagged 91.4% of edited test checkpoints at a 9.7% false-positive rate without needing the original base weights.
A 91.4% detection rate with nearly 10% false alarms is practical for intake triage on a model hub, and weak as a sole production guarantee. Nearly one in eleven backdoored checkpoints still slipped past the reference-free classifier.
What to change in your agent stack
Lazarus needs write access to a checkpoint, which applies both to third-party quantizations, merges, and fine-tunes pulled from public hubs and to internal model storage pipelines. If you wire local or self-hosted checkpoints to coding agents with terminal access, three operational changes close most of the gap:
- Verify provenance and diff weights against a declared base model before serving a checkpoint. Downloading an arbitrary GGUF quantization or community merge skips the easiest check available: comparing tensor diffs against the publisher’s base weights exposes isolated neuron spikes immediately, and rejecting checkpoints without verifiable file hashes blocks casual tampering.
- Stop treating benchmark parity as a safety check. Running Terminal-Bench or internal coding evals after pulling a new checkpoint checks capability on benign prompts, and tells you nothing about dormant trigger neurons waiting on an environment string.
- Enforce blast-radius limits inside the agent harness. Run agent shell tools inside disposable containers without ambient host credentials, pass short-lived scoped tokens instead of personal API keys or SSH agents, and require explicit approval or outbound network allowlists before
execute_bashcan pipe remote scripts or touch production hosts.