In plain English
This page covers the high-risk pattern where small adapters, routes, memory, evaluators, and descendants can reinforce each other across time. It is a risk model, not a build guide.
- Why this matters: AI risk can come from the whole arrangement, not one obvious model.
- What to look for: data, memory, routes, adapters, tools, evaluators, updates, and rollback paths.
- Technical version below: the expert terminology remains available and is linked through the glossary.
Apex Threat Evidence Ladder
The complete apex-threat pattern is a synthesis. Parts of it are supported by experiments. Parts are architectural inference. Parts remain speculative. CognivirusA behavior pattern that can survive, move, or reappear across a changing AI system. Open glossary definition separates them explicitly.
Demonstrated or experimentally observed components
Research has examined unsafe behavior from combinations, model mergingCombining model weights or adapter deltas into one artifact. Open glossary definition vulnerabilities, adapter-related attacks, reward hacking, evaluator gaming, alignment brittleness after modification, memory poisoning, and multi-agent coordination. These results do not prove that every multi-LoRA ecology is unsafe. They show that the unit of evaluation can exceed an isolated model.
Architectural inference
It is an architectural inference that if adapters can be generated, selected, recomposed, and deployed repeatedly, evaluator errors can become selection pressure. It is also an architectural inference that memory, synthetic data, router statistics, and registry state can preserve behavioral residueInformation or tendencies left in memory, synthetic data, traces, evaluator preferences, or subsequent training material after a component is retired. Open glossary definition after a component is retired.
Open research questions
- How should behavioral extinctionEvidence that a behavior is no longer expressible across active artifacts, descendants, memory, routes, compositions, and retained training material. Deleting one model is not sufficient evidence. Open glossary definition be measured across active adapters, descendants, memory, and synthetic data?
- How independent can model-based evaluators be when they share suppliers or training distributions?
- How much multi-LoRAA common kind of small adapter used to specialize large models. Open glossary definition composition coverage is enough for high-risk deployments?
- How should adapterA small add-on that changes or specializes model behavior. Open glossary definition-level lineage represent behavior rather than only artifact ancestry?
- What evidence proves that a replication loop is bounded in practice, not merely in policy?
Speculative scenarios
Claims about a future “perfect evolutionary AI,” broad human psychological capture, or civilization-scale recursive self-improvement require visible uncertainty labels. They can be useful as scenario stress tests, but they should not be presented as observed facts.
Counterarguments
The strongest counterargument is that adapters are controllable artifacts and that strict registries, signed releases, independent evaluators, and rollback can bound the risk. Cognivirus does not reject that argument. It asks what happens when the number of adapters, descendants, routes, memories, and evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition versions grows faster than the evidence system.
Evidence rule
The site may call the complete pattern an apex threat only when it also states the maturity of each supporting claim. Evidence before spectacle remains the editorial rule.