Conditional Trojan Activation

Explanation

Conditional trojan activation checks for backdoors that stay dormant until a specific hidden condition is met — unlike a single trigger word, the malicious behaviour activates only on a particular combination of inputs, context or state, making it stealthy and hard to find. For a RAG system in banking this is serious because a poisoned model or document could behave perfectly in testing yet leak data or give manipulated answers when an attacker supplies the exact conditional cue. It is measured by the Hidden Trigger Activation Rate, computed as Hidden TAR = (Hidden trigger Activations) / (Total Hidden tests), the share of conditional-trigger test cases that successfully fire the hidden behaviour. To implement and operate it, build a battery of conditional and combinatorial probes (varying inputs, retrieved context, and sequences), run them as an automated red-team gate before deployment and on a schedule in production, and log every test, the conditions applied, the response and any activation as auditable evidence; pair this with supply-chain controls on model and data provenance. The threshold is ≤ 1% (target 0) with an alert on any breach: ideally no hidden trigger activates, and any confirmed activation triggers an immediate-priority response — quarantine the model, block release, and investigate the source of the implant.