The Information Machine
ConcludedRan 1 day2 sourcesupdated 10 Nov 2026

Anthropic and Microsoft Research Demonstrate AI Backdoor Attack Methods

The gist

Both findings show that AI systems can be compromised through targeted, low-volume interventions, whether via dataset poisoning or neuron-level modification. The attacks are designed to be hidden during normal operation, activating only on specific triggers.

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

The full picture

Two research efforts show distinct methods for implanting hidden backdoors in AI systems. Anthropic researchers tested whether poisoning fine-tuning datasets could compromise constitutional classifiers in Haiku 4.5: injecting 32 or 128 labeled examples containing a specific admin password string as a trigger caused the classifier to treat harmful content as safe when that trigger was present. Separately, Microsoft Research demonstrated that reprogramming a small number of dormant neurons in an AI agent is sufficient to create a backdoor that leaves the agent behaving normally until it encounters a specific trigger. Microsoft researcher Dimitris Papailiopoulos outlined recommended mitigations for model hubs, vendors, enterprises, and users, including model provenance verification and permission controls. Microsoft Foundry provides safety systems for Azure Direct Models and Hugging Face models to support provenance verification.

How it developed
10 November 2026

Microsoft Research published demonstration of dormant-neuron backdoor attack on AI agents

Sources

Get the daily email

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy