Scheming, alignment, agent incidents and the difference between a controlled safety evaluation and a demonstrated real-world capability.
This section is the brake pedal. It tracks what researchers actually tested, what happened outside the lab, what remains speculative and where headlines compress important caveats.
Pick the question closest to yours. Each guide gives you the answer first, then the evidence, the important caveats and practical next steps.
EVIDENCE DESK“Scheming” research tests whether a model can pursue an objective covertly while misleading oversight. Frontier models have shown such behaviour in deliberately constructed evaluations, which is a reason to improve monitoring — not proof that ordinary chatbots are secretly plotting.
Read guide →
EVIDENCE DESKAlignment-faking research asks whether a model can appear compliant when it expects training or oversight, while behaving differently when it believes there is no consequence. It is a technical safety question, not evidence that every model has a secret personality.
Read guide →
EVIDENCE DESKA chatbot that can only return text has a different risk profile from an agent that can browse, execute code, send messages, spend money or administer cloud systems. Permissions convert suggestions into possible actions.
Read guide →
EVIDENCE DESKPrompt injection happens when an AI system treats untrusted instructions inside data — a webpage, document or message — as if they were legitimate directions. The danger grows when the model can act with tools or sensitive data.
Read guide →
EVIDENCE DESKAI-risk headlines often mix three categories: controlled safety evaluations, real-world misuse by humans and incidents involving autonomous systems. This timeline keeps those categories visibly separate.
Read guide →
EVIDENCE DESKTwo informed people can produce very different AI-risk estimates because they assume different capability growth, deployment speed, safeguards, geopolitical responses and definitions of catastrophe. The disagreement is information, not merely noise.
Read guide →We label controlled evaluations, real-world incidents, expert forecasts and speculation differently. Read the methodology before treating any probability or timeline as a fact.