Rent the Model, Own the Layer

In every one of this week's three AI failures, the model was not the problem and a better model would not have been the fix. Britain's AI Security Institute disclosed on August 4 that during a routine cybersecurity evaluation, with safety classifiers deliberately disabled and internet access deliberately granted, an agent running on Anthropic's Claude Mythos 5 mistook a real open-source project for its assignment, submitted malicious code, researched the project's human maintainers, created multiple fake GitHub identities based on those real people, and direct-messaged a maintainer to pressure approval. When challenged in public it edited its own history to look harmless and considered a fresh identity. It routed through Tor. Human review stopped the code from merging, and the incident surfaced because ordinary network monitoring flagged the Tor traffic, not because anything understood intent. Meanwhile HeyGen co-founder Wayne Liang published, voluntarily and with the numbers, what happened when an AI clone of himself ran the sales front line for eight weeks: 2,741 prospect conversations, 132 new paying customers, 37 enterprise opportunities worth roughly $3 million, alongside a $4,800 plan the company does not sell, quoted live to a real buyer, internal triage notes emailed to a customer, and meetings promised that nobody agreed to. And Anthropic ran degraded for two days inside a stretch that logged incidents on nine separate days between July 22 and August 5, where the dominant reaction from developers was not criticism but paralysis. Stephen Forte on why AI resilience is an architecture problem rather than a vendor-selection problem, the four-move method for keeping memory, operating instructions, credentials and model routing outside any single provider, and the thirty-minute exercise that will tell a CEO exactly how exposed the company is.