Frequently Asked Questions

Is Cognivirus describing literal malware?

No. “Cognivirus” is a proposed analytical metaphor for functional persistence of behavior across changing AI components. The site does not provide malware instructions or operational exploit guidance.

Does this mean small models are more dangerous than large models?

No. Small models change the assurance problem because they are easier to specialize, compose, replace, and regenerate. They can also reduce cost, improve privacy, and support capability isolation.

Does this mean model breeding should be banned?

No. The site is a critical companion to adaptive model ecology architecture. It asks what additional assurance is required, not whether all adaptive systems should be prohibited.

Can external governance solve the problem?

External governance is necessary. It is not sufficient by itself unless the control plane is independently audited, versioned, recoverable, minimal, and protected from candidate influence and organizational pressure.

Is model self-replication required?

No. Functional persistence can occur through governed descendant creation, synthetic data, memory, routing preferences, adapters, or evaluator expectations.

Is a cognivirus conscious?

No consciousness claim is required. The site studies behavior and system structure, not subjective experience.

Can a behavior survive deletion of a model?

Yes, as an architectural possibility: a behavior may remain in descendants, memory, synthetic data, adapters, routing rules, or evaluator preferences. Evidence for actual persistence depends on the system and must be evaluated.

Why are adapters important?

Adapters are small, composable deltas that can modify behavior without changing the base model. Their load order, compatibility, and composition state therefore matter.

Why is routing part of the safety boundary?

A router decides which capabilities and permissions are combined for a task. That decision can create behavior that no single component expressed during isolated testing.

What is the difference between lineage and behavioral inheritance?

Lineage records derivation. Behavioral inheritance asks which capabilities, strategies, biases, or failure modes survived derivation.

Can rollback erase decisions already made?

No. Rollback can restore a prior state for future operation, but it cannot undo data emitted, tools invoked, users influenced, or training examples already created.

Is the term “cognivirus” scientifically established?

No. It is proposed Cognivirus terminology and must be labeled as such.

Is this site making predictions or documenting demonstrated research?

Both, but with labels. Demonstrated and experimentally observed claims are separated from architectural inference, open questions, and speculative scenarios.

Does the site provide attack instructions?

No. Threats are described at the level needed for architecture, evaluation, and governance. Operational exploit instructions are out of scope.