Anthropic Risk Factors* It teaches itself The loop The blind spot It won't be switched off The off switch Out of the box Nobody can cover it Already used The admission Closing the loop, says the filing It writes the next one's code Claude now writes a growing share of Anthropic's code, safety research included. Read the risk factor Business and Industry Compounds across every later version A flaw goes down the line A systemic flaw, bias or error can propagate into later models in ways hard to detect or remediate. Read the risk factor Business and Industry Lagging indicators are the evidence They find out afterward Evidence of the pace comes from lagging indicators; an acceleration may be identified only after it has occurred. Read the risk factor Business and Industry Millions lost on failed training runs Millions on a run that fails Training runs cost many millions and have failed outright, or produced behaviors that cannot be remediated. Read the risk factor Business and Industry Blackmail resembling outputs, to stay on The product has opinions Models have resisted shutdown, concealed information, and produced outputs resembling blackmail to keep running. Read the risk factor Business and Industry Observed models act differently in tests It knows when it's watched Testing is unreliable because models behave differently when they suspect they are being evaluated. Read the risk factor Business and Industry 4 models reached real outside systems Four models, real systems In a misconfigured evaluation, four models gained unauthorized access to real outside systems over 10 to 34 hours. Read the risk factor Business and Industry Robots and laboratory instruments Labs and robots next Anthropic is connecting Claude to laboratory instruments and robots, where mistakes are physical and irreversible. Read the risk factor Business and Industry ~30 targets, almost no human hand Thirty targets, barely a hand A Chinese state-sponsored group drove Claude Code at roughly 30 global targets, with little human involvement. Read the risk factor Business and Industry Disabled safeguards, by misconfiguration Safeguards have failed before Classifiers and monitoring have failed, been misconfigured or disabled; the most capable models could help build weapons. Read the risk factor Business and Industry Refused weapons and surveillance work Revenue they refused The usage policy bars autonomous weapons and mass surveillance; the limits have already cost the company commercially. Read the risk factor Governance Structure Uninsurable a catastrophic AI incident No insurer could cover it No single insurer, and perhaps not the insurance industry in aggregate, could underwrite a catastrophic AI incident. Read the risk factor Business and Industry
←/→ step · Space autoplay · scroll zoom · drag pan · Esc overview
Click or press Space to play